Sei sulla pagina 1di 347

UNIVE RSITA' DI PADO VA

FA COLTA ' DI INGE GNERIA

SCUOLA DI DOTTORA TO DI RICE RCA IN


ING EGNE RIA INDUSTRIA LE

INDIRIZZO ENE RGETICA

CICLO : XX IV

Aer od yn am ic shape
o p tim izatio n o f ro tary win g
air craft com p onents using
advan ced m ultiobjective
evolutio nar y algor ithm s

DIRETTORE DE LLA S CUOLA :


Prof. A lbe rto Mi rand ola

SUPE RV ISORE :
Prof. E rne sto Be nini

DO TTORANDO: Clau dio Comis Da Ronco

3 1 gen naio 2 012


A chi mi vuole bene
Sommario

Lo scopo di questa tesi di Dottorato in Energetica, finanziata da AgustaWest-


land, consiste nella progettazione e sviluppo di una procedura di ottimizzazione
multi-obiettivo, che comprende l’applicazione del GeDEA-II, un algoritmo geneti-
co/evolutivo recentemente sviluppato dall’autore presso l’Università di Padova.
Tale algoritmo permette di effettuare analisi di ottimizzazione multi-obiettivo,
sfruttando l’approccio del tutto generale che va sotto il nome di “Ricerca del Fronte
di Pareto”. Rispetto ad altri algoritmi evolutivi multi-obiettivo “state-of-the-art”,
esso presenta operatori di crossover e mutazione innovativi, che ne migliorano in
maniera significativa le performance.
Questo ottimizzatore è accoppiato con i codici CFD commerciali e gratuiti,
rispettivamente Fluent® and OpenFOAM®. Il pacchetto Altair Hyperworks, codice
ufficiale presso AgustaWestland per il pre-processing di fusoliere di elicottero, è
scelto quale software per la parametrizzazione free-form.
I casi test scelti per dimostrare l’efficacia di tale procedura consistono nell’ot-
timizzazione aerodinamica della regione frontale del tilt rotor dimostrativo ERICA,
e della presa d’aria ]1 dell’elicottero AW101. Tali casi costituiscono problemi sti-
molanti sia da un punto di vista puramente ingegneristico, sia da un punto di
vista industriale. A partire dall’elaborazione della geometria, e procedendo con
la discussione dei risultati ottenuti, ogni passo della procedura di ottimizzazione
è descritto in dettaglio, con particolare enfasi dedicata al ciclo di ottimizzazione,
sviluppato dall’autore sia in ambiente UNIX, sia in ambiente Windows.
I risultati ottenuti dimostrano l’efficacia dell’approccio di ottimizzazione basato
sull’algoritmo GeDEAII, scelto per sviluppare questo lavoro.
Inoltre, vengono forniti alcuni consigli inerenti lo sviluppo futuro di questa
procedura di ottimizzazione, con l’obiettivo di migliorare ulteriormente le capacità
e la robustezza del ciclo di ottimizzazione.

i
Abstract

The aim of this Doctoral Thesis, sponsored by AgustaWestland, is the design


and development of a multi-objective optimization procedure that involves the
application of the GeDEA-II, a powerful and time-saving evolutionary algorithm
recently developed by the author at the University of Padova, able to perform
multi-objective optimization analyses with the general approach of the Pareto fron-
tier search. When compared to other state-of-the-art multi-objective evolutionary
algorithms, it features novel crossover and mutation operators, and demonstrated
superior performance.
This optimizer supervises an automatic optimization loop involving the CFD
commercial and free, open source solvers, respectively Fluent® and OpenFOAM®.
Altair Hyperworks package is chosen as the free-form-deformation parameteriza-
tion engine.
The test cases chosen to demonstrate the strength of the procedure implemented
concern the aerodynamic optimization of the AgustaWestland ERICA nose region,
and the optimization of the intake ]1 of the AW101 helicopter, that is really
challenging problems from both the engineering and the industrial point of view.
Starting from the the geometry elaboration and proceeding to the results
discussion, each step of the optimization procedure is described in details, with
particular focus on the automatic optimization loop, directly programmed by the
author in both UNIX/Linux and Windows environments.
The results obtained surely demonstrate the effectiveness of the multi-objective
approach chosen to carry out this work.
Furthermore, some suggestions for future improvements and developments are
provided, with the purpose to increase the strength of the discussed multi-objective
optimization tool.

iii
Why research?

When I first started my PhD three years ago I must admit I was a bit skeptical
about my future. I don’t know what I will do in my future but I certainly believe
that professional, international experience of a high level is essential to be able to
complete my accademic studies. In any case, three years later as a result this PhD
course has given me much personal satisfaction.
Why embark on a PhD research programme? Why do research? Well I believe
what makes human beings different from animals isn’t either their soul or their
intelligence but what makes man unique is his knowledge and the desire to push
himself ahead beyond his limits.
Giacomo Leopardi wrote between 1829 and 1830 “Night Song of a Wandering
Shepherd in Asia”, and at one point he says:

And when I gaze upon the stars at night,


In thought I ask myself,
“Why all these torches bright?
What mean these depths of air,
This vast, this silent sky,
This nightly solitude? And what am I?”

Gazing towards the sky the poet contemplates on the meaning of life and looks
for an answer to his question : “what is life”?
I think that researchers should draw inspiration from Leopardi’s poems. “Striv-
ing for the infinite” is one of the main trains of thought of the Romantic poets and
this is what should distinguish every researcher: push yourself and go beyond
your limits.
So what could be more wonderful and inspiring than doing research? Of
course it means making a lot of sacrifices and getting discouraged at times, but is
there anything more humane than doing research? I don’t think so, I believe it’s
important not to stifle one’s ambitions but keep the flame alive.
In conclusion I urge whoever reads this Thesis to reflect upon my words.
The future of research programmes in Italy at the moment is undergoing some
dark moments. Without Universities and Research programmes hypothetically
speaking we would all become like a bunch of trained circus animals. I think a
great country should be made up of well-educated people with strong critical
and thinking skills in order to make the right decisions for themselves about their
future.

v
vi

This country must invest in research to keep up the standard of research


programmes: you can’t put a price on knowledge and learning.
As far as I’m concerned I intend carrying on with this belief and heaven help
those who want to prevent me from pursuing my dreams!
Long-live Research! Long-live Italy!
Contents

Sommario i

Abstract iii

Why research? v

List of Figures xx

List of Tables xxii

Acronyms xxiii

Introduction xxvii

Chapter 1 1

1 Optimization techniques and Evolutionary algorithms 1


1.1 Optimization strategies: An overview . . . . . . . . . . . . . . . . . . 2
1.2 Direct search methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 The Simplex Algorithm . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Gradient-based Methods for Unconstrained Problems . . . . . . . . . 6
1.3.1 The Quasi-Newton Method . . . . . . . . . . . . . . . . . . . . 8
1.4 Gradient-based Methods for Constrained Problems . . . . . . . . . . 9
1.4.1 Sequential Quadratic Programming . . . . . . . . . . . . . . . 10
1.5 Evolutionary Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.1 Framework of an Evolutionary Algorithm . . . . . . . . . . . 12
1.6 Multi-Objective Optimization . . . . . . . . . . . . . . . . . . . . . . . 13
1.6.1 Pareto concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6.2 Traditional Solution Methods . . . . . . . . . . . . . . . . . . . 15
1.6.3 Multi-Objective Evolutionary Algorithms . . . . . . . . . . . . 16

Chapter 2 19

2 GeDEAII: A powerful and robust MOEA 19


2.1 Multiobjective Evolutionary Algorithms and the Problem of Diver-
sity Preservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

vii
viii CONTENTS

2.2 Genetic Diversity Evolutionary Algorithm


(GeDEA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Genetic Diversity Evolutionary Algorithm-II
(GeDEA-II) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1 The SIMPLEX crossover . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2 The Shrink mutation . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.3 Diversity preservation . . . . . . . . . . . . . . . . . . . . . . . 30
2.3.4 The Self-tuning operator . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Comparison with Other Multiobjective Evolutionary Algorithms . . 32
2.4.1 Test functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.3 Metric of Performance . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.4 Results of Comparison . . . . . . . . . . . . . . . . . . . . . . . 39
2.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

Chapter 3 53

3 Parameterization techniques for 3D shape optimization in aerodynamics 53


3.1 Shape parameterization techniques: An Overview . . . . . . . . . . . 53
3.2 The HyperMorph Parameterization Tool . . . . . . . . . . . . . . . . 54
3.2.1 Domains-Handles Approach . . . . . . . . . . . . . . . . . . . 55
3.2.2 Morph-Volume Approach . . . . . . . . . . . . . . . . . . . . . 55
3.2.3 Shapes as Design Variables . . . . . . . . . . . . . . . . . . . . 56

Chapter 4 57

4 Introduction to CFD 57
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2.1 The Continuity Equation . . . . . . . . . . . . . . . . . . . . . 58
4.2.2 The Momentum Equation . . . . . . . . . . . . . . . . . . . . . 60
4.2.3 The Energy Equation . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3 Turbulence Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.1 The RANS Technique . . . . . . . . . . . . . . . . . . . . . . . 68
4.3.2 The κ − ω turbulence model . . . . . . . . . . . . . . . . . . . 73
4.4 Generic form of the Governing Equations for CFD . . . . . . . . . . . 75
4.5 The Finite Volume method: an overview . . . . . . . . . . . . . . . . . 76

Chapter 5 81

5 A Survey of the Techniques for Drag Reduction 81


5.1 The Tilt-Rotor Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2 Power requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2.1 Fuselage and body induced drag breakdown . . . . . . . . . 85
5.3 Overview of aircraft drag sources . . . . . . . . . . . . . . . . . . . . . 86
5.4 Skin friction drag reduction . . . . . . . . . . . . . . . . . . . . . . . . 87
CONTENTS ix

5.4.1 Laminar Flow Control (LFC) . . . . . . . . . . . . . . . . . . . 87


5.4.2 Natural Laminar Flow (NLF) . . . . . . . . . . . . . . . . . . . 90
5.4.3 Hybrid Laminar Flow Control (HLFC) . . . . . . . . . . . . . 91
5.4.4 Riblets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.4.5 LEBUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.5 Pressure drag reduction . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.5.1 Synthetic Jet actuators . . . . . . . . . . . . . . . . . . . . . . . 94
5.5.2 Vortex Generators . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.5.3 Active Vortex Generators . . . . . . . . . . . . . . . . . . . . . 100
5.5.4 Micro Vortex Generators . . . . . . . . . . . . . . . . . . . . . . 101
5.6 Interference drag reduction . . . . . . . . . . . . . . . . . . . . . . . . 102
5.6.1 Gaster bumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.6.2 Slot suction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.7 Lift-induced drag reduction . . . . . . . . . . . . . . . . . . . . . . . . 103
5.7.1 Winglets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.8 Miscellaneous drag sources . . . . . . . . . . . . . . . . . . . . . . . . 104
5.9 Regions of application . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

Chapter 6 107

6 Introduction to Intake Aerodynamics 107


6.1 Preliminar considerations . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2 Subsonic inlets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.3 Inlet Operating Conditions . . . . . . . . . . . . . . . . . . . . . . . . 110
6.3.1 Internal flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.3.2 External flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.4 Intake-Engine Integration: Flow Distortion Issue . . . . . . . . . . . . 115
6.4.1 Total pressure distortion issue and Definition of Intake Per-
formance Parameters . . . . . . . . . . . . . . . . . . . . . . . . 116
6.5 Effect of Distortion on Compressor Performance . . . . . . . . . . . . 118

Chapter 7 121

7 Results and Discussion 123

I Test case A:

Aerodynamic shape optimization of the AW101 air intake ]1 125


7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.2 Optimization Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.3 The object of the optimization . . . . . . . . . . . . . . . . . . . . . . . 129
7.4 Geometry elaboration . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.5 Mesh generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.5.1 Superficial mesh . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.5.2 Volume grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
x CONTENTS

7.6 Fluid-dynamic model set up . . . . . . . . . . . . . . . . . . . . . . . . 136


7.7 Boundary and operating conditions . . . . . . . . . . . . . . . . . . . 139
7.8 Grid sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.8.1 Aerodynamic characterization of the intake baseline geometry145
7.9 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.9.1 Morphing Constraints . . . . . . . . . . . . . . . . . . . . . . . 150
7.9.2 Design Variables Definition . . . . . . . . . . . . . . . . . . . . 150
7.10 Formulation of the intake optimization problem . . . . . . . . . . . . 156
7.11 Discussion of results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.12 Reverse Engineering Procedure in CATIA® V5 . . . . . . . . . . . . . 162

II Test case B:

Aerodynamic shape optimization of the frontal region of the ERICA


tiltrotor using Ansys Fluent® as the CFD solver 167
7.13 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.14 Optimization Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.15 The object of the optimization . . . . . . . . . . . . . . . . . . . . . . . 170
7.16 Geometry elaboration . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.17 Mesh generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.17.1 Superficial mesh . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.17.2 Volume grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.18 Fluid-dynamic model set up . . . . . . . . . . . . . . . . . . . . . . . . 179
7.18.1 Boundary and operating conditions . . . . . . . . . . . . . . . 180
7.19 Grid sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.20 Validation of the CFD model . . . . . . . . . . . . . . . . . . . . . . . 185
7.21 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.21.1 Morphing Constraints . . . . . . . . . . . . . . . . . . . . . . . 189
7.21.2 Design Variables Definition . . . . . . . . . . . . . . . . . . . . 190
7.22 D.O.E. and Student analysis . . . . . . . . . . . . . . . . . . . . . . . . 191
7.23 Optimization study: discussion of results . . . . . . . . . . . . . . . . 198
7.24 Reverse Engineering Procedure in CATIA® V5 . . . . . . . . . . . . . 205

III Test case C:

Aerodynamic shape optimization of the frontal region of the ERICA


tiltrotor using OpenFOAM® as the CFD solver 209
7.25 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
7.26 Optimization Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 211
7.27 The object of the optimization . . . . . . . . . . . . . . . . . . . . . . . 213
7.28 Mesh generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.28.1 Superficial mesh . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.28.2 Volume grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
CONTENTS xi

7.29 Fluid-dynamic model set up . . . . . . . . . . . . . . . . . . . . . . . . 221


7.29.1 Boundary and operating conditions . . . . . . . . . . . . . . . 222
7.30 Fluid dynamic characterization of the baseline geometry . . . . . . . 224
7.31 Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
7.31.1 Morphing Constraints . . . . . . . . . . . . . . . . . . . . . . . 230
7.31.2 Design Variables Definition . . . . . . . . . . . . . . . . . . . . 231
7.32 Formulation of the intake optimization problem . . . . . . . . . . . . 235
7.33 Discussion of results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236

Conclusions 247

Appendices 253

A DC (60) computation Procedure 253

B Methodology for assessing compliance with the visibility requirements 255

C Automatic loop optimization files 259


C.1 GeDEAII.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
C.2 Command-function.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
C.3 Command.tcl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
C.4 Tgrid-batch.jou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
C.5 Fluent-batch.jou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
C.6 OpenFOAM-autorun.sh . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
C.7 Student.m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279

D A technical overview of the open-source CFD package OpenFOAM® 283


D.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
D.2 OpenFOAM framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
D.3 OpenFOAM accessibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
D.4 Model setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
D.5 0 directory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
D.5.1 Variable p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
D.5.2 Variable T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
D.5.3 Variable U . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
D.5.4 Variable Turbulence Kinetic Energy κ . . . . . . . . . . . . . . 292
D.5.5 Variable Specific Dissipation Rate ω . . . . . . . . . . . . . . . 293
D.6 Constant directory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
D.6.1 PolyMesh directory . . . . . . . . . . . . . . . . . . . . . . . . 294
D.6.2 Files turbulenceProperties and RASProperties . . . . . . . . . . . 295
D.6.3 File thermophysicalProperties . . . . . . . . . . . . . . . . . . . . 296
D.7 system directory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
D.7.1 File controlDict . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
D.7.2 File decomposeParDict . . . . . . . . . . . . . . . . . . . . . . . . 299
D.7.3 File fvSchemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
D.7.4 File fvSolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
xii CONTENTS

Acknowledgements 307

Bibliography 309
List of Figures

1.1 Flowchart of Search Algorithm. . . . . . . . . . . . . . . . . . . . . . . 1


1.2 The original simplex, the reflection of one vertex through the cen-
troid of the opposite face, and the resulting reflection simplex. . . . 6
1.3 Crossover and mutation children. . . . . . . . . . . . . . . . . . . . . . 13
1.4 Typical framework of an Evolutionary Algorithm. . . . . . . . . . . . 14
1.5 MOOP evaluation mapping. . . . . . . . . . . . . . . . . . . . . . . . . 15

2.1 Approximate Pareto-optimal set reached by GeDEAII, GeDEA, IBEA,


NSGA-II and SPEA2 provided with the same evolutionary operators,
on ZDT1 test function. . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 The reflection step of the simplex algorithm applied to a problem in
R2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Application of Simplex crossover in a multi objective context. . . . . 27
2.4 Approximate Pareto-optimal set reached by IBEA, NSGA-II and
SPEA2 provided with and without the SimplexCrossover, on ZDT6
(on the left) and DTLZ6 (on the right) test functions. . . . . . . . . . . 28
2.5 Approximate Pareto-optimal sets reached by IBEA, NSGA-II and
SPEA2 provided with the ShrinkMutation operator (on the left side
of each subfigure) and with the PolynomialMutation operator (on the
right side of each subfigure). . . . . . . . . . . . . . . . . . . . . . . . 29
2.6 Significance of the Hypervolume indicator as far as convergence (a,
AT THE TOP), and diversity (b, AT THE BOTTOM) is concerned. . . 39
2.7 Test function T1 (convex). . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.8 Test function T2 (non-convex). . . . . . . . . . . . . . . . . . . . . . . . 41
2.9 Test function T3 (discrete). . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.10 Test function T4 (multi-modal). . . . . . . . . . . . . . . . . . . . . . . 42
2.11 Test function T6 (non-uniform). . . . . . . . . . . . . . . . . . . . . . . 43
2.12 Test function KUR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.13 Box plots based on the Hypervolume metric. Each square contains
six box plots representing the distribution of Hypervolume values for
the six algorithms. Results refer to the ZDT test suite. . . . . . . . . . 44
2.14 Test function DTLZ1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.15 Test function DTLZ2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.16 Test function DTLZ4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.17 Test function DTLZ3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

xiii
xiv LIST OF FIGURES

2.18 Test function DTLZ5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48


2.19 Test function DTLZ7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.20 Test function DTLZ6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.21 Box plots based on the Hypervolume metric. Each square contains
six box plots representing the distribution of Hypervolume values for
the six algorithms. Results refer to the DTLZ test suite. . . . . . . . . 51

3.1 Handles-Domain technique. . . . . . . . . . . . . . . . . . . . . . . . . 55


3.2 Morph-Volumes application technique. . . . . . . . . . . . . . . . . . 56

4.1 An infinitesimal fluid control volume - The Continuity equation. . . 59


4.2 An infinitesimal fluid control volume - The Momentum Equation. . 60
4.3 An infinitesimal fluid control volume. The Work contribution to the
Energy equation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.4 An infinitesimal fluid control volume. The Heat contribution to the
Energy equation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.6 Skin friction coefficient of NACA 23012 airfoil. XFOIL output . . . . 67
4.7 Laminar to turbulent induced transition. . . . . . . . . . . . . . . . . 68
4.8 Direct Numerical Simulation of two reacting flows (image available
at website http://www.ecs.umass.edu/mie/faculty/debk/Research/
dns.html). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.9 Volume rendered image of a Large Eddy Simulation of a non-
premixed swirl flame (taken from [86]). . . . . . . . . . . . . . . . . . 69
4.10 Viscous wake impinging a stator vane row of an aero-engine LP
turbine. (Axial velocity perturbation magnitude coloured)(taken
from [20]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.11 A representation of structured and unstructured mesh for the finite-
volume method (full symbols denote element vertices and open
symbols at the center of the control volumes denote computational
nodes). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.12 A comparison between a fully tetrahedral unstructured grid (ON
THE LEFT) and a viscous hybrid mesh (ON THE RIGHT). . . . . . . 78
4.13 Control volume for the two-dimensional continuity equation prob-
lem. Image taken from [92] . . . . . . . . . . . . . . . . . . . . . . . . 79

5.1 An artistic impression of the ERICA concept (taken from [37]). . . . 82


5.2 Different ERICA possible configurations (taken from [66]). . . . . . . 82
5.3 ERICA tiltable wing (taken from [37]). . . . . . . . . . . . . . . . . . . 83
5.4 Contributions of different drag sources for a typical transport aircraft
(taken from [1]). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.5 Transition mechanism on swept wings (taken from [85]). . . . . . . . 88
5.6 LFC system designed by McDonnell- Co. (taken from [48]). . . . . . 89
5.7 NLF, LFC and HLFC concepts for wing (taken from [48]). . . . . . . 91
5.8 Sketch of riblet geometry (taken from [95]). . . . . . . . . . . . . . . . 92
5.9 Sketch of a tandem arrangement of a LEBU device (taken from [35]). 94
LIST OF FIGURES xv

5.10 The vortex wake behind an upswept afterbody (taken from [89]). . . 95
5.11 Schematic of a synthetic jet (taken from [6]). . . . . . . . . . . . . . . 96
5.12 Iso surfaces of the instantaneous vorticity magnitude. (a) Uncon-
trolled case; (b) controlled case. (taken from [98]) . . . . . . . . . . . 97
5.13 Model mounted in the wind tunnel: detail of vortex generators
(taken from [85]). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.14 Different shapes of vortex generators (taken from [42]). . . . . . . . . 100
5.15 Schematic of the proposed separation control actuation concept
(taken from [97]). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.16 Winglet attached to wing tip (taken from [47]). . . . . . . . . . . . . . 103

6.1 CAD model of the ERICA intake and exhaust system. . . . . . . . . 109
6.2 Schematic drawing of subsonic diffuser geometry with straight
centerline (taken from [28].) . . . . . . . . . . . . . . . . . . . . . . . . 110
6.3 Schematic drawing of an S-duct with two pockets of swirling flows
(known as the secondary flow patter) generated by the bends (taken
from [28].) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.4 Typical stream line patterns in subsonic inlets (taken from [43].) . . . 111
6.5 Possible locations of boundary layer separation (taken from [43].) . . 112
6.6 One-dimensional model for external deceleration study(taken from
[28].) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.7 Minimum frontal area ratio for various values of c Pmax (s = 0.5 in
equation 6.1)(taken from [43].) . . . . . . . . . . . . . . . . . . . . . . 114
6.8 Different types of total pressure distortion (taken from [28].) . . . . . 117
6.9 Illustration of total-pressure contours and ϑ sector for definition of
distortion coefficient (taken from [79].) . . . . . . . . . . . . . . . . . . 118
6.10 Effect of inlet total pressure distortion on compressor stability (taken
from [28].) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.11 Effects of circumferential inlet distortion on multistage axial com-
pressor performance (taken from [77].) . . . . . . . . . . . . . . . . . 120
6.12 Effect of circumferential distortion sector angle on surge pressure
ratio (taken from [77].) . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.13 Effect of number of sectors, on surge pressure ratio (taken from [77].)121
6.14 Effect of the extent of circumferential spoiling on compressor perfor-
mance (taken from [28].) . . . . . . . . . . . . . . . . . . . . . . . . . . 121

7.1 Flow chart of the complete optimization procedure. . . . . . . . . . . 129


7.2 CATIA® V5 models of AW101 model without main and tail rotors. . 130
7.3 Virtual wind-tunnel layout with the AW101 left air intake system
installed on a virtual flat plate. . . . . . . . . . . . . . . . . . . . . . . 131
7.4 Close-up of the CAD model of the isolated air intake]1 . . . . . . . . 131
7.5 A detailed view of the Air Intake and the Aerodynamic Interface
Plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.6 A detailed view of the Air Intake with the dummy duct and the
compressor face. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
xvi LIST OF FIGURES

7.7 Superficial mesh over the Intake system. . . . . . . . . . . . . . . . . 135


7.8 Superficial mesh over the virtual wind tunnel walls. . . . . . . . . . . 135
7.9 Longitudinal view of the whole volumetric mesh on a center line
plane. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.10 Close-up of the volume mesh near the intake system. . . . . . . . . . 137
7.11 Close-up of the volume mesh: boundary layer near the compressor
face. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.12 AIP DC (60) sensitivity to dummy duct length. Forward Flight con-
dition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.13 AIP DC (60) sensitivity to dummy duct length. Hovering Flight
condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.14 Contours of wall y+ over the Air Intake walls: forward flight (on the
left) and hovering (on the right). . . . . . . . . . . . . . . . . . . . . . 145
7.15 Contours of total pressure ([Pa]) over the air intake walls: forward
flight (on the left) and hovering (on the right). . . . . . . . . . . . . . 146
7.16 Streamlines pattern inside the air intake: forward flight (on the left)
and hovering (on the right). . . . . . . . . . . . . . . . . . . . . . . . . 147
7.17 Total pressure distribution ([Pa]) over the AIP surface: forward flight
(on the left) and hovering (on the right). . . . . . . . . . . . . . . . . . 147
7.18 Circumferential distribution of DC (60) for the intake ]1 baseline
model: forward flight. . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.19 Circumferential distribution of DC (60) for the intake ]1 baseline
model: hovering flight. . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.20 Location of the DC (60) worst sector for the intake ]1 baseline model:
forward flight (on the left) and hover conditions (on the right). . . . 148
7.21 Geometrical constraints on the air intake surfaces: front view (on
the left) and top view (on the right). . . . . . . . . . . . . . . . . . . . 150
7.22 Cut planes used for the shape parameterization. . . . . . . . . . . . . 152
7.23 Parametric shape Sh1, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.24 Parametric shape Sh2, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.25 Parametric shape Sh3, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.26 Parametric shape Sh4, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.27 Parametric shape Sh5, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.28 Parametric shape Sh6, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.29 Parametric shape Sh7, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
7.30 Parametric shape Sh8, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
LIST OF FIGURES xvii

7.31 Parametric shape Sh9, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.32 The entire set of geometries calculated during the optimization run. 158
7.33 Final Pareto front after 20 generations. . . . . . . . . . . . . . . . . . . 159
7.34 Optimized solution geometrical configuration: comparison with the
baseline intake geometry. . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.35 Contours of total pressure ([Pa]) over a series of transversal sections
along the intake duct in hovering condition: comparison of the
baseline (on the left) and optimized (on the right) solutions. . . . . . 161
7.36 Contours of total pressure ([Pa]) over a series of transversal sections
along the intake duct in forward flight condition: comparison of the
baseline (on the left) and optimized (on the right) solutions. . . . . . 162
7.37 Contours of total pressure ([Pa]) over a series of transversal sections
along the intake duct in hovering condition: comparison of the
baseline (on the left) and optimized (on the right) solutions. . . . . . 163
7.38 Contours of total pressure ([Pa]) over a series of transversal sections
along the intake duct in forward flight conditions: comparison of
the baseline (on the left) and optimized (on the right) solutions. . . . 163
7.39 Streamlines pattern inside the air intake in hover: comparison of
baseline (on the left) and optimized (on the right) configurations. . . 164
7.40 Streamlines pattern inside the air intake in forward flight: compari-
son of baseline (on the left) and optimized (on the right) configurations.164
7.41 Comparison of baseline and optimized DC60 distribution in hovering.164
7.42 Comparison of baseline and optimized DC60 distribution in forward
flight. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.43 The final optimized intake geometry. . . . . . . . . . . . . . . . . . . 166
7.44 Superimposition of the baseline and the optimized intake CAD model.166
7.45 Flow chart of the complete optimization procedure. . . . . . . . . . . 171
7.46 ERICA nose geometry (taken from [74]) . . . . . . . . . . . . . . . . . 172
7.47 ERICA nose (on the right) and cylindrical portion of the fuselage
(on the left) (taken from [74]) . . . . . . . . . . . . . . . . . . . . . . . 172
7.48 ERICA tail cone (taken from [74]) . . . . . . . . . . . . . . . . . . . . . 172
7.49 Virtual wind-tunnel layout with the tiltrotor fuselage. . . . . . . . . . 173
7.50 Superficial mesh over ERICA nose. . . . . . . . . . . . . . . . . . . . . 175
7.51 Superficial mesh over the virtual wind tunnel walls. . . . . . . . . . . 176
7.52 Histograms representing ERICA surface mesh quality statistics:
Aspect-Ratio (on the left); Skewness (on the right). . . . . . . . . . . . . 177
7.53 Longitudinal view of the whole volumetric mesh around the tiltrotor
fuselage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.54 Close-up of the volume mesh near the fuselage. . . . . . . . . . . . . 179
7.55 Close-up of the volume mesh: boundary layer over the canopy. . . . 179
7.56 Nose drag sensitivity to cross-sectional area. . . . . . . . . . . . . . . 183
7.57 Nose drag sensitivity to wind-tunnel box length. . . . . . . . . . . . 184
7.58 Contours of wall y+ over the ERICA fuselage. . . . . . . . . . . . . . 186
xviii LIST OF FIGURES

7.59 Model-scaled tiltrotor fuselage drag curve: simulations vs. experi-


ments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.60 Fixed nodes over the fuselage and tail surfaces. . . . . . . . . . . . . 190
7.61 Shape sh1-sh2 definition. . . . . . . . . . . . . . . . . . . . . . . . . . . 192
7.62 Shape sh3 definition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
7.63 Shape sh4 definition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.64 Shape sh5 definition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.65 Shape sh6-sh7 definition. . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.66 Shape sh8-sh9 definition. . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.67 Shape sh10-sh11 definition. . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.68 Shape sh12-sh13 definition. . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.69 Shape sh14-sh15 definition. . . . . . . . . . . . . . . . . . . . . . . . . . 195
7.70 Student analysis on the design variables of ERICA nose total drag. . 197
7.71 Convergence history of the GeDEAII driven optimization. . . . . . . 199
7.72 Geometrical comparison between the baseline and the optimized
nose configurations: longitudinal view (top) and top view (bottom). 200
7.73 Pressure coefficient contours over the baseline configuration (left)
and the optimized configuration (right). . . . . . . . . . . . . . . . . . 201
7.74 Skin friction coefficient contours over baseline configuration (left)
and the optimized configuration (right). . . . . . . . . . . . . . . . . . 201
7.75 Visualization of the planes selected for pressure and viscous drag
analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
7.76 Pressure drag component over the selected planes: comparison of
baseline and optimized configurations. . . . . . . . . . . . . . . . . . 203
7.77 Viscous drag component over the selected planes: comparison of
baseline and optimized configurations. . . . . . . . . . . . . . . . . . 204
7.78 The reverse engineering procedure: cyan curves represent the group
1, while black curves represent the group 2. . . . . . . . . . . . . . . . 206
7.79 The point introduced during the curve creation in order to respect
the original variation of the curvature. . . . . . . . . . . . . . . . . . . 207
7.80 The cockpit construction lines. . . . . . . . . . . . . . . . . . . . . . . 207
7.81 The final optimized nose geometry. . . . . . . . . . . . . . . . . . . . 208
7.82 Superimposition of the baseline and the optimized CAD models. . . 208
7.83 Flow chart of the complete optimization procedure accommodating
OpenFOAM as the CFD colver. . . . . . . . . . . . . . . . . . . . . . . 212
7.84 ERICA nose geometry in CATIA V5. . . . . . . . . . . . . . . . . . . . 213
7.85 ERICA windscreen geometry in CATIA V5. . . . . . . . . . . . . . . . 214
7.86 ERICA fuselage geometry in CATIA V5. . . . . . . . . . . . . . . . . . 214
7.87 ERICA wf-fearing geometry in CATIA V5. . . . . . . . . . . . . . . . . 215
7.88 ERICA wing-sx geometry in CATIA V5. . . . . . . . . . . . . . . . . . 215
7.89 ERICA sponson-sx geometry in CATIA V5. . . . . . . . . . . . . . . . 215
7.90 ERICA tail geometry in CATIA V5. . . . . . . . . . . . . . . . . . . . . 216
7.91 Wind tunnel cad model. . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.92 Superficial mesh model: wind tunnel component. . . . . . . . . . . . 217
7.93 Superficial mesh model: nose component. . . . . . . . . . . . . . . . . 217
LIST OF FIGURES xix

7.94 Histograms representing ERICA surface mesh quality statistics: Skew-


ness (on the left); Aspect-Ratio (on the right). . . . . . . . . . . . . . . . 219
7.95 Longitudinal view of the whole volumetric mesh around the tiltrotor
fuselage. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
7.96 Close-up of the volume mesh near the fuselage. . . . . . . . . . . . . 220
7.97 Close-up of the volume mesh: boundary layer over the canopy. . . . 221
7.98 First order residuals with low under relaxation factors. . . . . . . . . 224
7.99 First-second order residuals with high under relaxation factors. . . . 225
7.100Convergence behaviour of CL. . . . . . . . . . . . . . . . . . . . . . . 225
7.101Convergence behaviour of CD. . . . . . . . . . . . . . . . . . . . . . . 226
7.102Convergence behaviour of CM. . . . . . . . . . . . . . . . . . . . . . . 226
7.103Pressure coefficient contours C p for the baseline geometry. . . . . . . 227
7.104Skin friction coefficient C f for the baseline case. . . . . . . . . . . . . . 227
7.105Streamlines along a longitudinal plane. Baseline geometry. . . . . . . 227
7.106Streamlines in the nose region. Baseline geometry. . . . . . . . . . . . 228
7.107Friction lines over the nose for the baseline geometry. . . . . . . . . . . 228
7.108Contours of wall y+ over the ERICA fuselage. . . . . . . . . . . . . . 229
7.109Fixed (shown in red) and free (shown in green) surfaces. . . . . . . . 230
7.110Cut planes used for the shape parameterization. . . . . . . . . . . . . 232
7.111Parametric shape Sh1, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
7.112Parametric shape Sh2, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
7.113Parametric shape Sh3, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
7.114Parametric shape Sh4, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
7.115Parametric shape Sh5, applied to the intake model with scaling
factor α = +1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
7.116Comparison between the baseline (black line) and optimized (red
line) configurations. Sh1 shape. . . . . . . . . . . . . . . . . . . . . . . 236
7.117Comparison between the baseline (black line) and optimized (red
line) configurations. Sh2 shape. . . . . . . . . . . . . . . . . . . . . . . 237
7.118Comparison between the baseline (black line) and optimized (red
line) configurations. Sh3 shape. . . . . . . . . . . . . . . . . . . . . . . 237
7.119omparison between the baseline (black line) and optimized (red line)
configurations. Sh5 shape. . . . . . . . . . . . . . . . . . . . . . . . . . 237
7.120Rectilinear graph. Comparison between baseline e optimized configu-
rations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
7.121Comparison of Cp between baseline (UP) and optimized (DOWN)
configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
7.122Comparison of Cf between baseline (UP) and optimized (DOWN)
configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
7.123Visualization of the cut planes selected for pressure and viscous
drag analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
xx LIST OF FIGURES

7.124Pressure profile over Plane A. . . . . . . . . . . . . . . . . . . . . . . . 241


7.125Pressure profile over Plane B. . . . . . . . . . . . . . . . . . . . . . . . 241
7.126Pressure profile over Plane C. . . . . . . . . . . . . . . . . . . . . . . . 242
7.127Pressure profile over Plane D. . . . . . . . . . . . . . . . . . . . . . . . 242
7.128Pressure profile over Plane E. . . . . . . . . . . . . . . . . . . . . . . . 243
7.129τ profile along plane A. . . . . . . . . . . . . . . . . . . . . . . . . . . 243
7.130τ profile along plane B. . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
7.131τ profile along plane C. . . . . . . . . . . . . . . . . . . . . . . . . . . 244
7.132τ profile along plane D. . . . . . . . . . . . . . . . . . . . . . . . . . . 245
7.133τ profile along plane E. . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

A.1 View of AIP from Downstream Showing 60°Sector for DC (60) Cal-
culation at Angle. (taken from [37]) . . . . . . . . . . . . . . . . . . . 254

B.1 Total vision envelope of the baseline ERICA configuration. . . . . . . 255


B.2 Assumed azimuth angles convention. . . . . . . . . . . . . . . . . . . 256
B.3 Assumed elevation angles convention. . . . . . . . . . . . . . . . . . . 257

C.1 Flow chart of the GeDEAII code. . . . . . . . . . . . . . . . . . . . . . 260

D.1 Typical framework of an OpenFOAM case. . . . . . . . . . . . . . . . 287


List of Tables

1.1 Main differences between Classical and Genetic Algorithms. . . . . . 11

2.1 Settings of the Self-tuning operator. . . . . . . . . . . . . . . . . . . . 32


2.2 Original and proposed number of generations for the ZDT and
DTLZ test suites. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5.1 Power required breakdown. . . . . . . . . . . . . . . . . . . . . . . . . 85


5.2 Fuselage and body-induced drag breakdown. (NICETRIP wind
tunnel tests, 2008. [67]) . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3 Tilt-rotor regions suitable for application of the mentioned devices. . 105

7.1 Flight Conditions for Air Intake optimization. . . . . . . . . . . . . . 131


7.2 Finally selected number of elements and target element size for each
fuselage component. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.3 Ansys Tgrid® settings for volume mesh generation. . . . . . . . . . . 137
7.4 Air properties for the intake optimization. . . . . . . . . . . . . . . . 139
7.5 Solver settings and discretization schemes. . . . . . . . . . . . . . . . 139
7.6 Pressure inlet specification at inlet. . . . . . . . . . . . . . . . . . . . . 140
7.7 Pressure-outlet specification at outlet. . . . . . . . . . . . . . . . . . . 140
7.8 Pressure-outlet specification at compressor face. . . . . . . . . . . . . 141
7.9 Total pressure on AIP sensitivity to dummy duct length. . . . . . . . 142
7.10 Superficial mesh features of the coarse and refined model. . . . . . . 144
7.11 Total pressure on AIP sensitivity to mesh refinement. . . . . . . . . . 144
7.12 Trial intake optimization results: design parameters values for the
selected optimized individual. . . . . . . . . . . . . . . . . . . . . . . 158
7.13 Trial intake optimization results: Total pressure drop and DC60
reduction with respect to the baseline geometry . . . . . . . . . . . . 161
7.14 Flight Conditions for Nose optimization (taken from [74]) . . . . . . 173
7.15 Finally selected number of elements and target element size for each
fuselage component . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.16 Quality parameters limits specified for ERICA surface mesh. . . . . 176
7.17 Mesh statistic for the finally selected ERICA superficial grid. . . . . 177
7.18 Ansys Tgrid® settings for volume mesh generation. . . . . . . . . . . 178
7.19 Air properties for ERICA nose optimization. . . . . . . . . . . . . . . 180
7.20 Solver settings and discretization schemes. . . . . . . . . . . . . . . . 180

xxi
xxii LIST OF TABLES

7.21 Pressure inlet specification at inlet. . . . . . . . . . . . . . . . . . . . . 181


7.22 Pressure-outlet specification at outlet. . . . . . . . . . . . . . . . . . . 182
7.23 Drag sensitivity to wind tunnel cross-sectional area dimensions. . . 183
7.24 Drag sensitivity to wind-tunnel box length. . . . . . . . . . . . . . . . 184
7.25 Drag sensitivity to superficial mesh size. . . . . . . . . . . . . . . . . 184
7.26 Volume mesh refinement influence on the nose drag characteristics. 185
7.27 Flowfield characteristics at wind tunnel conditions. . . . . . . . . . . 187
7.28 Boundary conditions on the fluid domain for the validation case. . . 187
7.29 Model-scaled tiltrotor fuselage drag coefficient: simulations vs. ex-
periments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.30 Design variables deformation ranges. . . . . . . . . . . . . . . . . . . 198
7.31 Objective function values for the baseline and optimized nose con-
figurations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
7.32 Absolute and normalized design variables values of the optimized
configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
7.33 x-coordinate of the generated planes normal to x-axis. . . . . . . . . 205
7.34 Quality parameters limits specified for ERICA surface mesh. . . . . 218
7.35 Mesh statistic for the finally selected ERICA superficial grid. . . . . 218
7.36 Ansys Tgrid® settings for volume mesh generation. . . . . . . . . . . 219
7.37 Air properties for ERICA nose optimization. . . . . . . . . . . . . . . 222
7.38 1st order discretization schemes and solver settings. . . . . . . . . . . 223
7.39 1st − 2nd order discretization schemes and solver settings. . . . . . . 223
7.40 Operating conditions and OpenFOAM boundary conditions settings.224
7.41 Boundary conditions set onto fuselage walls. . . . . . . . . . . . . . . 224
7.42 Aerodynamic forces acting onto the entire baseline model. . . . . . . 225
7.43 CL, CD and CM coefficients for the baseline geometry. . . . . . . . . 226
7.44 CD comparison and drag force [N] between the baseline and the
optimized configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . 238
7.45 Design Variables and shapes displacements . . . . . . . . . . . . . . . 238

B.1 Up and down angles as a function of azimuth angles for the ERICA
baseline configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

D.1 Discretization schemes available in ddtSchemes. . . . . . . . . . . . . . 300


D.2 Interpolation schemes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
D.3 Discretization schemes available in gradSchemes. . . . . . . . . . . . . 301
D.4 Divergence schemes available in divSchemes. . . . . . . . . . . . . . . 303
D.5 Surface normal gradient schemes available in snGradSchemes. . . . . 303
D.6 Limiter coefficients available in limited schemes of snGradSchemes. . 303
D.7 Surface normal schemes available in laplacianSchemes. . . . . . . . . . 304
Acronyms

AIP: Aerodynamic Interface Plane.


DC (60): Distortion Coefficient calculated assuming a 60 ° angular sector.
D.O.E.: Design Of Experiments.
EP: Evolutionary Programming.
ES: Evolution Strategies.
FUSELAGE INCIDENCE: Angle between Fuselage Water Line and the free-stream
wind direction [Positive nose-up].
GeDEAII: Genetic Diversity Evolutionary Algorithm II.
GeDEM: Genetic Diversity Evaluation Method.
HPC: High Performance Computing.
MOEA: Multi Objective Evolutionary Algorithm.
MOOP: Multi Objective Optimization Problem.
OAT: Outside Air Temperature.
SQP: Sequential Quadratic Programming.

xxiii
Introduction

Aviation is an essential element of today’s global society, bringing people and


cultures together and creating economic growth. The air transport industry is
paying a lot of attention to growing public concern about the environmental issues
of air pollution, noise and climate change. Although today air transport only
produces 2% of man-made CO2 emissions, this is expected to increase to 3% by
2050.
Clean Sky1 is a Joint Technology Initiative (JTI) that will develop breakthrough
technologies to reduce aviation environmental impact. The Clean Sky JTI will be
one of the largest European research projects ever, with a budget estimated at 1.6
billion euros, equally shared between the European Commission and industry, over
the period 2008 - 2013. This public-private partnership will speed up technological
breakthrough developments and shorten the time to market for new solutions
tested on Full Scale Demonstrators. The accelerated research process that Clean
Sky offers represents an unprecedented opportunity for rapid progress in the
introduction of green technology into aviation. The Clean Sky JTI gathers 86
organization over 16 countries:
• 54 industries, including AgustaWestland S.p.A, which, as leader of the Eu-
ropean helicopter market, is involved in the Green Rotocraft ITD (Integrated
Technology Demonstrator) project;
• 15 research centers;
• 17 Universities, including University of Padova as AgustaWestland partner.
One of the goals of the Clean Sky JTI is the 50% reduction of CO2 emission
through drastic reduction of aircraft fuel consumption, that means, from the
aeronautic industries point of view, to adapt the aircraft and rotorcraft design
procedures to new severe standards about aerodynamic drag and engines efficiency.
Organized around six major platforms or Integrated Technology Demonstrators
(ITDs), the Clean Sky project will lead to the development of in-flight or ground
demonstrators. Among these ITDs, the Green Rotorcraft platform is specifically
devoted to helicopters and tilt-rotor aircrafts, whose operations are expected to
grow sharply in the future to meet the increasing demands of European citizens:
• Medical service for safe and quick transport of patients and living organs
needed for transplantation;
1 Please visit the web site http://www.cleansky.eu for more elucidations.

xxv
xxvi Introduction

• Passenger transport from city heliports to airports, and also between cities or
areas where an efficient surface transport network cannot be developed for
geographical or economical reasons;
• Search, rescue, police and border patrol missions.
In turn, within Green Rotorcraft ITD, six technological projects have been
defined and divided between the two leaders AgustaWestland and Eurocopter.
The first project (GRC1) focuses on blades, where the optimization of shapes
and the use of active control systems are intended to reduce noise and consumption.
The second project (GRC2) concerns airframe design, which must be made more
aerodynamic for more efficient power use in flight. Depending on the rotorcraft
configuration and weight class the aerodynamic cumulated drag benefits are
currently estimated in the range 10 − 15% which would translate into a 4 to 5%
fuel consumption reduction for the same payload and mission.
The third project (GRC3) covers the integration of innovative electrical systems
that will eliminate engine air bleeds and the need for hydraulic fluid.
The fourth project (GRC4) examines the integration of diesel engines on light
helicopters and will eventually lead to the production and flight testing of a
demonstrator.
The fifth (GRC5) is devoted to aeroacoustics research.
The sixth (GRC6) but by no means least important project tackles eco-design,
an area that is of equal interest to airplane manufacturers. Nevertheless, the Eco-
Design ITD also offers certain specificities for helicopter manufacturers, such as
the surface treatment of dynamic components in the power transmission system.
These six projects will be followed by the synthesis work that will technologically
assess the solutions after integration.
The subject of this Thesis deals with the second Green Rotorcraft project
(GRC2), since the content of the research illustrated here aims at gaining a deeper
insight into the computer-based optimization procedure of helicopters and tilt
rotor airframe components.
In this context, AgustaWestland and the University of Padova are focusing
their attentions on the opportunity to integrate the classical trial and error design
approach with a more advanced approach, based on the automatic search of
optimal solutions using optimization algorithms.
The present Thesis becomes part of the AgustaWestland/University of Padova
aerodynamic optimization research program, with the purpose to build up an
automatic optimization procedure able to deal with mono and multi-objective
aerodynamic, but even multi-disciplinary, optimization problems, in order to create
an optimization tool able to drastically improve the aerodynamic characteristics of
fuselage component under investigation. The AgustaWestland requirements for
this optimization tool are the following:
• It has to include the softwares at the moment available at AgustaWestland, re-
garding CAD geometry elaboration (CATIA), mesh generation (HyperMesh)
and parametrization (HyperMorph), and advanced aerodynamic analysis
calculation (Fluent® and/or OpenFOAM®);
Introduction xxvii

• It has to be able to run also on high computational power machines, like the
UNIX/Linux based cluster computers2 ;

• It has to enable the user to treat not only single-objective, but also multi-
objective optimization problems.

The last item of the previous list, enforces the choice of an intrinsically multi-
objective optimization algorithm, like the home-made algorithm developed during
my PhD course, i.e. the GeDEA-II, as optimization engine. The central point of the
present thesis is, therefore, to implement the proper interface among the optimiza-
tion engine GeDEAII, the aerodynamic solvers (Fluent® and/or OpenFOAM®),
and parameterization tool Hypermorph, all of them capable of running on both
Windows-based workstations and Linux/UNIX-based clusters.
To prove the applicability of that optimization procedure on real engineering
problems, three test problems were addressed.
The aerodynamic optimization of the nose region, belonging to the AgustaWest-
land new tilt-rotor concept ERICA [37], has been carried out, demonstrating that
such kind of parametric analysis may become a really useful design tool in the
future. To perform this work, both Fluent® and OpenFOAM® CFD solvers where
exploited. Related test cases are Test Case B and Test Case C, respectively, presented
in Chapter 7.
Moreover, the aerodynamic optimization of the AW101 left air intake system
has been performed, in order to prove further the versatility and robustness of the
optimization loop, and presented within Test case A of Chapter 7.
The Thesis is structured as follows.
Chapter 1 provides all the theory knowledge about optimization methods.
GeDEAII algorithm is presented in Chapter 2, with a particular emphasis on its
framework, its operators and its performance measured against some state-of-the-
art algorithms, on state-of-the-art benchmark problems.
Morphing technologies are introduced in Chapter 3.
A general introduction to CFD is then presented in Chapter 4.
Chapter 5 gives general informations about aircraft external aerodynamics,
along with the most promising techniques aiming at reducing aircraft overall drag.
Chapter 6 constitutes a brief introduction to the intake aerodynamics, in order
to prepare the ground for the optimization problem regarding the AW101 left
air-intake system discussed in Part I of Chapter 7.
Chapter 7 represents the main segment of the Thesis, where the optimization
procedure and its application to the AW101 left intake and ERICA nose region are
wholly described.
Finally the conclusions which can be drawn from the whole work are presented,
along with some suggestions for future works and developments.
2A computer cluster is a group of linked computers, working together closely so that in many
respects they form a single computer. The components of a cluster are commonly, but not always,
connected to each other through fast local area networks. Clusters are usually deployed to improve
performance and/or availability over that of a single computer, while typically being much more
cost-effective than single computers of comparable speed or availability
xxviii Introduction
Chapter 1

Optimization techniques and


Evolutionary algorithms Optimization Methods: Advanced Topics in Optimization - Direct and Indirect Search 1
Methods

Module – 8 Lecture Notes – 4

Direct and Indirect Search Methods


Most of the real world system models involve nonlinear optimization with com-
Introduction
plicated objective functions or constraints for which analytical solutions (solutions
using quadratic programming, geometric programming, etc.) are not available.
In such cases one of the possible solutions is the search algorithm in which, the
objective function is first computed with a trial solution and then the solution is
sequentially improved based on the corresponding objective function value till
convergence. A generalized flowchart of the search algorithm in solving a nonlinear
optimization with decision variable X i , is presented in Figure 1.1. i, is presented in
Fig. 1.

Start with Trial


Solution Xi, Set i=1

Compute Objective
function f(Xi)

Generate new
solution Xi+1

Compute Objective
function f(Xi+1)
Set i=i+1

Convergence Check
No
Yes
Optimal Solution Fig. 1 Flowchart of Search Algorithm
Xopt=Xi

D Nagesh Kumar,
Figure IISc, Bangaloreof
1.1: Flowchart Search Algorithm. M8L4

1
Most of the real world system models involve nonlinear optimization with complicated
objective functions or constraints for which analytical solutions (solutions using quadratic
programming, geometric programming, etc.) are not available. In such cases one of the
possible solutions is the search algorithm in which, the objective function is first computed
with a trial solution and then the solution is sequentially improved based on the
corresponding objective function value till convergence. A generalized flowchart of the
search algorithm in solving a nonlinear optimization with decision variable X
2 1. Optimization techniques and Evolutionary algorithms

Optimization methods can be basically classified into three main categories,


based on the approach followed to reach the optimal solution:

1. Direct or derivative-free methods;

2. Classical, gradient-based, methods;

3. Stochastic methods.

A direct search algorithm for numerical search optimization depends on the


objective function only through ranking a countable set of function values. It
does not involve the partial derivatives of the function and hence it is also called
non-gradient or zeroth order method.
Gradient-based algorithms, also called the descent methods, depend on the first
(first-order methods) and often second derivatives (second-order methods) of the
objective function.
Stochastic methods are optimization algorithms which incorporate probabilistic
(random) elements, either in the problem data (the objective function, the con-
straints, etc.), or in the algorithm itself (through random parameter values, random
choices, etc.), or in both [83]. The concept contrasts with the deterministic opti-
mization methods, where the values of the objective function are assumed to be
exact, and the computation is completely determined by the values sampled so far.
Stochastic optimization methods include:

• Evolutionary algorithms [39];

• Simulated annealing algorithms [51];

• Particle Swarm algorithms [50];

In this chapter a short introduction to general optimization problems formula-


tion is provided, including a discussion about the Direct methods. Solving strategies
for both unconstrained and constrained optimizations using some classical, gradient
based, techniques are then briefly introduced while the description of stochastic
methods is focused on genetic algorithms, since it is the method chosen to perform
the aerodynamic shaoe optimizations presented in Chapter 7. In the last part of
the chapter an introduction to multi-objective optimization theory is also provided,
with particular focus on the Genetic Diversity Evolutionary Method (GeDEM) that is
implemented into the multi-objective optimization code Genetic Diversity Evolution-
ary AlgorithmII (GeDEAII) used to perform the work described in the following
chapters.

1.1 Optimization strategies: An overview


Optimization techniques are used to find a set of design parameters x = {x1 ,
x2 ,. . . xn }, that can in some way optimize a given function relative to a design
problem. In a simple case this may be the minimization or maximization of some
system characteristic that is dependent on x. In a more advanced formulation, the
1.2 Direct search methods 3

optimization problem of the objective function f (x), to be minimized or maximized,


can be stated as in Equation 1.1 (n representing the number of decision variables):

min f ( x ) (1.1)
x ∈Rn

subject to the constraints in the form of:

• equality constraints, Gi (x) = 0 ( i = 1,. . . ,me );

• inequality constraints, Gi (x)≤ 0 (me +1,. . . ,m);

• parameter bounds like xl ≤ 0 x ≤ xu , where xl is the parameters lower bound


and xl is the parameters upper bound respectively.

where x is the vector of the design variables, x ∈ Rn , f(x) is the objective function
that returns a scalar value, in mono-objective optimization cases f ( x ) : Rn → R,
or a vector of p elements where p is the number of objectives in the multi-objective
problem considered, f(x):Rn → R p ), and the vector function G(x) returns the values
of the equality and inequality constraints evaluated at the actual value of the design
variables x, G(x):Rn → Rm . An efficient and accurate solution to this problem is
not only dependent on the size of the problem in term of design variables and
constraints, but also on characteristics of the objective and constraint functions.
When both the objective function and the constraints are linear function of the
design parameters the problem is know as a Linear Programming problem (LP).
Quadratic Programming (QP) concerns the minimization or maximization of a
quadratic objective function that is linearly constrained. For both these kinds of
problems, reliable solution procedures are readily available. More complicated
to solve is the Non-linear Programming problem (NP) where both the objective
function and constraints are non-linear function of the design parameters. Generally
a solution of the non-linear problem requires an iterative procedure to establish a
search direction at each iteration; this is usually achieved by the solution of an LP,
a QP, or an unconstrained subproblem.

1.2 Direct search methods


A lot of problems derived from real applications are characterized by an
unknown analytical expression of both the objective function and constraints.
This situation occurs, for example, when complex physic systems are described,
analyzed and controlled optimizing the results of computer based simulations, and
the objective function and/or constraints values are evaluated by post-processing
the results of these simulations. This is also the case of the application studied in
this Thesis. It is clear that in these situations the gradient based approach may not
to be suitable, therefore it is of practical interest the development of optimization
methods which do not require derivative informations. Direct search or Derivative-
free is the name of a category of optimization methods which respond to this
requirement. As opposed to the more traditional methods discussed from section
4 1. Optimization techniques and Evolutionary algorithms

References, which use informations about the gradient and higher derivatives to
search for an optimal solution, a direct search algorithm searches a set of points
around the current point , looking for one where the value of the objective function
is lower than the value at the current point. Direct search methods are suitable to
solve problems for which the objective function is not differentiable, stochastic, or
even discontinuous.
Some of the direct search algorithms for solving nonlinear optimization, which
requires objective functions, are described below:

1. Random Search Method: This method generates trial solutions for the op-
timization model using random number generators for the decision vari-
ables. Random search method includes random jump method, random walk
method and random walk method with direction exploitation. Random jump
method generates huge number of data points for the decision variable
assuming a uniform distribution for them and finds out the best solution
by comparing the corresponding objective function values. Random walk
method generates trial solution with sequential improvements which is gov-
erned by a scalar step length and a unit random vector. The random walk
method with direct exploitation is an improved version of random walk
method, in which, first the successful direction of generating trial solutions is
found out and then maximum possible steps are taken along this successful
direction.

2. Grid Search Method: This methodology involves setting up of grids in the


decision space and evaluating the values of the objective function at each
grid point. The point which corresponds to the best value of the objective
function is considered to be the optimum solution. A major drawback of this
methodology is that the number of grid points increases exponentially with
the number of decision variables, which makes the method computationally
costlier.

3. Univariate Method: This procedure involves generation of trial solutions


for one decision variable at a time, keeping all the others fixed. Thus the
best solution for a decision variable keeping others constant can be obtained.
After completion of the process with all the decision variables, the algorithm
is repeated till convergence.

4. Pattern Directions: In univariate method the search direction is along the


direction of coordinate axis which makes the rate of convergence very slow.
To overcome this drawback, the method of pattern direction is used, in which,
the search is performed not along the direction of the coordinate axes but
along the direction towards the best solution. This can be achieved with Hooke
and Jeeves’method or Powell’s method. In the Hooke and Jeeves’-method [46],
a sequential technique is used consisting of two moves: exploratory move and
the pattern move. Exploratory move is used to explore the local behavior
of the objective function, and the pattern move is used to take advantage
1.2 Direct search methods 5

of the pattern direction. Powell’s method [69] is a direct search method


with conjugate gradient, which minimizes the quadratic function in a finite
number of steps. Since a general nonlinear function can be approximated
reasonably well by a quadratic function, conjugate gradient minimizes the
computational time to convergence.

5. Rosen Brock’s Method of Rotating Coordinates: This method [76] is a mod-


ified version of Hooke and Jeeves’method, in which, the coordinate system is
rotated in such a way that the first axis always orients to the locally estimated
direction of the best solution and all the axes are made mutually orthogonal
and normal to the first one.

6. Simplex Method: Simplex method is a conventional direct search algorithm


where the best solution lies on the vertices of a geometric figure in N-
dimensional space made of a set of N+1 points. The method compares the
objective function values at the N+1 vertices and moves towards the optimum
point iteratively. The movement of the simplex algorithm is achieved by
reflection, contraction and expansion.

Due to its semplicity, easy-to-use, and because of its popularity, the latter
method is briefly introduced in the next section.

1.2.1 The Simplex Algorithm


Simplex search methods are characterized by the simple device that they use
to guide the search. The first of the simplex methods is due to Spendley, Hext,
and Himsworth [84] in a paper that appeared in 1962. They were motivated
by the fact that earlier direct search methods required anywhere from 2n to 2n
objective evaluations to complete the search for improvement on the iterate. Their
observation was that it should take no more than n + 1 values of the objective
to identify a downhill (or uphill) direction. This makes sense, since n + 1 points
in the graph of f(x) determine a plane, and n + 1 values of f(x) would be needed
to estimate f(x) via finite-differences. At the same time, n + 1 points determine a
simplex. This leads to the basic idea of simplex search: construct a non-degenerate
simplex in Rn and use the simplex to drive the search. A simplex is a set of n
+ 1 points in Rn . Thus one has a triangle in R2 , and tetrahedron in R3 , etc. A
non-degenerate simplex is one for which the set of edges adjacent to any vertex
in the simplex forms a basis for the space. In other words, we want to be sure
that any point in the domain of the search can be constructed by taking linear
combinations of the edges adjacent to any given vertex. Not only does the simplex
provide a frugal design for sampling the space, it has the added feature that if one
replaces a vertex by reflecting it through the centroid of the opposite face, then the
result is also a simplex, as shown in Figure 1.2 a).
Once an initial simplex is constructed, the single move specified in the original
Spendley, Hext, and Himsworth simplex algorithm is that of reflection. This move
first identifies the “worst” vertex in the simplex (i.e., the one with the least desirable
The first of the simplex methods is due to Spendley, Hext, and Himsworth [21] in a paper that appeared
in 1962. They were motivated by the fact that earlier direct search methods required anywhere from 2n to
2n objective evaluations to complete the search for improvement on the iterate. Their observation was that
it should take no more than n + 1 values of the objective to identify a downhill (or uphill) direction. This
makes sense, since n + 1 points in the graph of f (x) determine a plane, and n + 1 values of f (x) would be
needed to estimate ∇f6(x) via finite-differences. At the same
1. Optimization techniques
time, n + 1 points determine a simplex.and
This Evolutionary algorithms
leads to the basic idea of simplex search: construct a nondegenerate simplex in Rn and use the simplex to
drive the search.
A simplex is a set of n + 1 points in Rn . Thus one has a triangle in R2 , and tetrahedron in R3 , etc.
A nondegenerate simplex is one for which the set of edges adjacent to any vertex in the simplex forms a
basis for the space. In other words, we want to be sure that any point in the domain of the search can be
constructed by taking linear combinations of the edges adjacent to any given vertex.
Not only does the simplex provide a frugal design for sampling the space, it has the added feature that
if one replaces a vertex by reflecting it through the centroid of the opposite face, then the result is also
xk proceed
a simplex, as shown in Figure 3.1. This, too, is a frugal feature because it means that one can xk
parsimoniously, reflecting one vertex at a time, in the search for an optimizer.

0 r 0

Fig. 3.1. The original simplex, the reflectionFig.


of one vertex
3.3. Thethrough the centroid
original simplex,of the
withopposite face, and the
the reflection, resulting
expansion, and two possible contraction simplices, along with the shrink
reflection simplex. (a) Reflection move (b) Nelder and Mead’s additional moves
step toward the best vertex xk , when all else fails.
Once an initial simplex is constructed, the single move specified in the original Spendley, Hext, and
Figure 1.2: The original simplex, the reflection of one vertex
Himsworth simplex algorithm is that of reflection. This move first identifies the “worst” vertex in the
through the centroid of the
opposite the
face, reflection
and the point,
resulting whereas the
reflection contraction
simplex. steps allow
simplex (i.e., the one with the least desirable objective value) and then reflects the worst simplex through
for more conservative moves by halving the length
ofreflected
the centroid of the opposite face. If the the step from
vertex thethecentroid
is still to then
worst vertex, either
nextthe reflection
choose the “next point or the worst vertex. Furthermore, in addition
worst” vertex and repeat the process.to(A allowing these
quick review adaptations
of Figure within
3.1 should confirma that
single iteration,
if the reflected these new possibilities have repercussions for future
vertex is not better than the next worst vertex, then if the “worst” vertex is once again chosen for reflection,
iterations as they deform (or, as the rationale goes, adapt) the shape of the original simplex.
it will simply be reflected back to where
objective value,it started,
thatthus creating
is the onean with
infinite the
cycle.)highest value in the case of a minimization
The ultimate goals are either to replaceNelder and
the “best” Mead
vertex alsooneresolved
(i.e., the the desirable
with the most question of what
objective to do if none of the steps tried bring acceptable
problem, or the one with the smallest one in the case of a maximizing problem)
improvement
value) or to ascertain that the best vertex byforadding
is a candidate a shrink
a minimizer. Until step: when
then, the all else
algorithm keepsfails, reduce the lengths of the edges adjacent to the
moving the simplex byand then
flipping some reflects the
vertex (other thanworst
the best simplex through
vertex) through the
the centroid centroid
of the opposite of the opposite face. The
current best vertex by half, as is also illustrated in Figure 3.3.
face. new formed simplex is now subject to a new process.
The basic heuristic is straightforward inThe Nelder–Mead
the extreme: we move simplex algorithm
a “worse” vertex has enjoyed
in the general directionenduring popularity. Of all the direct search methods,
The
of the remaining vertices (as thebyNelder–Mead
contribution
represented simplex
ofofNelder
the centroid the algorithm
and
remaining Meadwith
vertices), is [64]
the
theone
wasmostto often
expectationturn
of found in numerical
simplex search software
into anpackages.
The original
paper by Nelder and Mead is a Science Citation Index classic, with several
optimization algorithm with additional moves designed to accelerate the search. In thousand references across the
particular, scientific literaturewell-understood
it was already in journals rangingthat fromthe
Acta Anaesthesiologica
reflection Scandinavica
move preserved theto Zhurnal Fizicheskio
8

Khimii.
original shape In fact,
of the there isregardless
simplex, an entire book fromdimension.
of the the chemicalWhatengineering
Neldercommunity
and Mead devoted to simplex search
for optimization [28].
proposed was to supplement the basic reflection move with additional options
So why bother
designed to accelerate thewith looking
search by any further? Why
deforming the not rely exclusively
simplex in a way on that
the Nelder-Mead
they simplex method
suggested ifwould
one is going
better to employ
adapt atodirect search method?
the features of theThe answer: there
objective is the outstanding
function. To this question regarding
end, they added what are known as expansion and internal and external contraction optimizers. When the
the robustness of the Nelder-Mead simplex method that has long troubled numerical
method works,
moves, as shown in Figureit can1.2work
b).very well indeed,
In addition tooften
these finding a solution in far
improvements, fewer evaluations
Nelder and of the objective
function than other direct search methods. But it can also
Mead also resolved the question of what to do if none of the steps tried bring fail. One can see this in the applications literature,
acceptablefairly early on, frequently
improvement by addingreported as no more
a shrink than
step: “slow”
when allconvergence.
else fails, A systematic
reduce the study of Nelder-Mead,
when applied to a suite of standard optimization test problems,
lengths of the edges adjacent to the current best vertex by half, as is also illustrated also reported occasional convergence to a
on the rightnonstationary
side of Figurepoint 1.2of the
b).function [24]; the one consistent observation to be made was that in these instances
the deformation of the simplex meant that the search direction (i.e., the direction defined along the worst
vertex toward the centroid of the remaining vertices) became numerically orthogonal to the gradient.
These observations about the behavior of Nelder-Mead in practice led to two, relatively recent, investiga-
tions. The first [13], strives to investigate what can be proven about the asymptotic behavior of Nelder-Mead.
The results show that in R1 , the algorithm is robust; under standard assumptions, convergence to a station-
1.3 Gradient-based Methods
ary point is guaranteed. Some general for Unconstrained
properties Problems
in higher dimensions can also be proven, but none that
guarantee global convergence for problems in higher dimensions.
This is not surprising in light of a second recent result by McKinnon [16]. He shows with several examples
The derivative-free, or on
that limits exist proving globalsearch
gradient-based algorithms
convergence are basedto on
for Nelder-Mead: wit,the
thederiva-
algorithm can fail on smooth
tives or gradients of the objective function. The gradient of a function in n-
10
1.3 Gradient-based Methods for Unconstrained Problems 7

dimensional space is given by:

∂ f /∂x1
 
 ∂ f /∂x 
 2
.
 
∇f =  (1.2)
 
.

 
.
 
 
∂ f /∂xn
Gradient-based search algorithms include:
1. Steepest Descent (Cauchy) Method: In this method, the search starts from
an initial trial point X1 , and iteratively moves along the steepest descent
directions until the optimum point is found. Although the method is straight-
forward, it is not applicable to the problems having multiple local optima. In
such cases the solution may get stuck at local optimum points.

2. Conjugate Gradient (Fletcher-Reeves) Method: The convergence technique


of the steepest descent method can be greatly improved by using the con-
cept of conjugate gradient [29] with the use of the property of quadratic
convergence.

3. Newton’s Method: Newton’s method is a very popular method which is


based on Taylor’s series expansion. The Taylor’s series expansion of a function
f (x) at x=xi is given by (indicating with T the transposed matrix in all of the
following equations):
1
f ( x ) = f ( xi ) + ∇ f iT ( x − xi ) + ( x − xi )T [ Ji ]( x − xi ) (1.3)
2
where [ Ji ] = [ J ]| xi , is the Hessian matrix of f evaluated at the point xi . Setting
the partial derivatives of Eq. 1.3, to zero, the minimum value of f (x) can be
obtained.
∂ f (x)
= 0 f or j = 1, . . . , N (1.4)
∂x j
From Eq. 1.3 and 1.4

∇ f = ∇ f i + [ Ji ]( x − xi ) = 0 (1.5)

Eq. 1.5 can be solved to obtain an improved solution xi + 1

xi+1 = xi − [ Ji ]−1 ∇ f i (1.6)

The procedure is repeated till convergence for finding out the optimal solu-
tion.

4. Marquardt Method: Marquardt method is a combination method of both the


steepest descent algorithm and Newton’s method, which has the advantages
of both the methods, movement of function value towards optimum point
and fast convergence rate. By modifying the diagonal elements of the Hessian
matrix iteratively, the optimum solution is obtained in this method.
8 1. Optimization techniques and Evolutionary algorithms

5. Quasi-Newton Method: Quasi-Newton methods are well-known algorithms


for finding maxima and minima of nonlinear functions. They are based on
Newton’s method, but they approximate the Hessian matrix, or its inverse,
in order to reduce the amount of computation per iteration. The Hessian
matrix is updated using the secant equation, a generalization of the secant
method for multidimensional problems. For completeness of informations,
the following subsection will be entirely dedicated to this powerful method.

1.3.1 The Quasi-Newton Method


This method builds up curvature information at each iteration to formulate a
quadratic model problem of the form:

1
min x T Hx + c T x + b (1.7)
x 2

where H, the Hessian matrix, is a positive definite matrix, c is a constant vector


and b is a constant. The optimal solution for the problem occurs when the partial
derivative of x go to zero:
∇ f x ∗ = Hx ∗ + c = 0 (1.8)
x∗ being the optimal solution point, which can be written as:

x ∗ = − H −1 c = 0 (1.9)

Unlike this method, Newton-type methods, opposed to Quasi-Newton methods,


calculate H directly and proceed in a direction of descent to locate the minimum
after a number of iterations. To avoid the direct evaluation of H, that requires a
grate amount of computation, Quasi-Newton methods use the observed behavior
of f (x) and ∇f (x) to build up a curvature information to make an approximation to
H using an appropriate update technique. Generally the updating formula given
by Broyden, Fletcher, GoldFarb, and Shanno (BFGS) is thought to be the most effective
for the use in general purpose methods. The BFGS formula is presented in Eq. 1.10:

q x q Tx HkT skT sk Hk
Hk+1 = Hk + − (1.10)
q Tx sk skT Hk sk

where:

s k = x k +1 − x k
q k = ∇ f ( x k +1 ) − ∇ f ( x k )

As starting point, H0 can be set to any symmetric positive definite matrix. To avoid
the inversion of the Hessian, one can use a formula that makes an approximation of
the inverse Hessian at each update; a well know procedure of this kind is the DFP
formula of Davidon, Fletcher and Powell. This uses the same formula as the BFGS
method (Eq.1.10) except that qk is substituted for sk . The gradient information
is either supplied through analytically calculated gradients, or derived by partial
1.4 Gradient-based Methods for Constrained Problems 9

derivatives using a numerical differentiation method via finite differences. This


involves perturbing each of the design variables in turn and calculating the rate
of change in the objective function. At each major iteration, k, a line search is
performed in the following direction:

d = − Hk−1 ∇ f ( xk ) (1.11)

with the purpose to update the solution to:

x k +1 = x k + α k d (1.12)

in the order to meet the minimization condition f(xk+1 ) f( xk ), where αk is the line
search step evaluated with a line search technique [101].

1.4 Gradient-based Methods for Constrained Problems


In constrained optimization the general aim is to transform the problem into
an easier subproblem that can be solved and used as the basis of an iterative
process. Today the most widely used methods are all based on the solution of the
Karush-Kuhn-Tucker equations (KKT) which are necessary conditions for optimality
for a constrained optimization problem. The description of the KKT conditions can
be started with some simple definitions:

Convex programming problem: Problem in which both the objective function


and the inequality constraints are convex and the equality constraints are
linear functions of the design variables; in this category we can find both the
LP and QP problem;

Feasible point: Point that satisfies all the constraints;

Active constraint: Inequality constraint that is characterized by a null value in a


feasible point. By definition all the equal constraints are active.

It can be demonstrated that, if a given problem is a so called convex pro-


gramming problem, the KKT conditions are both necessary and sufficient for a
global solution point. Referring to the general problem stated in Eq.1.1, the KKT
conditions can be written as follows:

m
∇ f ( x ∗ ) + ∑ λi∗ ∇ Gi ( x ∗ ) = 0
i =1
Gi ( x ∗ ) = 0(i = 1, . . . , me )
Gi ( x ∗ ) ≤ 0(i = me + 1, . . . , m) (1.13)
λi∗ >0
m
∑ λi∗ Gi (x∗ ) = 0
i =1
10 1. Optimization techniques and Evolutionary algorithms

The first equation describes a cancelling of the gradient between the objective
function and the active constraints at the solution point. For the gradient to
be cancelled, Lagrangian multipliers (λi = 1,. . .,m) are necessary to balance the
magnitude deviation of the objective function and the constraints gradients. Since
only active constraints are included in this cancelling operation, constraints that
are not active not active must not be included in this operation and be set with a
corresponding Lagrangian multiplier equal to zero. This is stated implicitly in the
last two equations in Eq. 1.13.
The solution of the KKT equations forms the basis to many nonlinear pro-
gramming algorithms. These algorithms attempt to compute directly the Lagrange
multipliers. Constrained Quasi-Newton methods guarantee super-linear conver-
gence by accumulating second order information regarding the KKT equations
using a Quasi-Newton updating procedure. These method are commonly referred
to as Sequential Quadratic Programming (SQP) methods since a QP subproblem
is solved at each major iteration. These methods are briefly introduced in the
following section.

1.4.1 Sequential Quadratic Programming


Given the general problem in Eq. 1.1, the principal idea of the SQP concept is
the formulation of a QP subproblem based on a quadratic approximation of the
Lagrangian function:
m
L( x, λ) = f ( x ) + ∑ λi Gi ( x ) (1.14)
i =1

The QP subproblem is obtained by linearizing the nonlinear constraints. It can be


written as follows:

1
minn d T Hk d + ( xk )T d
d∈< 2
∇ Gi ( xk ) d + G( xk ) = 0(i = 1, . . . , me )
T
(1.15)
∇ Gi ( xk ) d + G( xk ) ≤ 0(i = me + 1, . . . , m)
T

This subproblem can be solved using any QP algorithm and the solution is
used to form a new iterate:
x k +1 = x k + α k d (1.16)
The step length parameter αk is determined by a line search procedure so
that an appropriate decrease in a merit function is obtained. The matrix H k is
an approximation of the Hessian matrix of the Lagrangian function and it can
be updated using the BFGS method previously discussed in section 1.3.1. A
nonlinearly constrained problem can often be solved with fewer iteration then
an unconstrained problem using SQP; the reason for this is that, because of the
limit in the feasible area, the optimizer can make well-informed decision regarding
directions of search and step length.
1.5 Evolutionary Algorithms 11

1.5 Evolutionary Algorithms


The genetic algorithms are methods of solving both constrained and uncon-
strained optimization problems based on a biological evolutionary strategy, also
called natural selection.
These methods repeatedly modifies a population of individual solutions. At
each optimization step, the genetic algorithm selects and randomly modifies indi-
viduals from the current population with the purpose to generate the individuals
for the next generation. Over the successive generations the population “evolves”
toward an optimal solution.
The genetic algorithm uses three main rules at each optimization step to create
the new population from the previous one:

1. Selection rules select the individuals, called parents, which contribute to the
population at the next generation;

2. Crossover rules combine two parents to form the children for the next
generation;

3. Mutation rules apply random changes to individual parents to form chil-


dren.

Genetic algorithms are really effective in general applications because they can
be used to solve a variety of optimization problems which are not well suited for
the standard algorithms just discussed in sections 1.2, 1.3, 1.4, including problems
in which the objective function is discontinuous, non-differentiable, stochastic,
or highly non-linear, such as the CFD functions. The main differences between
genetic algorithms and classical, derivative based, optimization algorithms are
summarized in the Table 1.1.
Classical Algorithms Evolutionary Algorithms
They generate a single point at each They generate a population of points
iteration. The sequence of points ap- at each iteration. The generations se-
proaches an optimal solution quence converges toward an optimal
solution
They select the next point in the se- They select the next population by com-
quence by deterministic calculations putations which uses random number
generators
Suitable for continuous, differentiable Suitable for every kind of function
functions
Appropriate for local optimum search Appropriate for global optimum search
They are intrinsically single-objective The fact that they evolve a population,
algorithms rather than a single solution, makes
them intrinsically multi-objective algo-
rithms

Table 1.1: Main differences between Classical and Genetic Algorithms.

In the genetic algorithm practice is common the use of some biological terms,
12 1. Optimization techniques and Evolutionary algorithms

which are now introduced in the follows, for better clarify the meaning of the
terminology used thorough the whole text of this Thesis.

• Fitness Function: the fitness function is the function that have to be optimized;
another way to call the objective function just named from section 1.1 to 1.4.1;

• Individual: an individual is any point in the search space to which the fitness
function is applicable; it corresponds to a given set of decision variables;

• Gene: a gene is a vector component of an individual. It represents another


way to call a design variable;

• Score: the score of an individual is its fitness function value;

• Population: the population is an array of individuals. In a problem character-


ized by a population size of n individuals and a number of design variables
equal to m, the population is an n × n matrix;

• Generation: a generation is a single Evolutionary algorithm iteration;

• Parents and Children: to create a new generation the Evolutionary Algorithm


selects certain individuals in the current population as parents, and uses
them to create individuals in the next generation, called children. Thanks to a
selection procedure, the algorithm is more likely to select parents with better
fitness value;

• Diversity: it refers to the average distance between individuals in a popu-


lation. This quantity is essential for the performance of a genetic algorithm
because it enables the algorithm to search a larger region of the space. Differ-
ent diversity preservation mechanism have been built.

For more elucidations regarding EA terminology, the reader is referred to the


reference [19].

1.5.1 Framework of an Evolutionary Algorithm


The block scheme in Figure 1.4 summarizes how a typical Evolutionary algo-
rithm works. In this scheme, the three methods used to generate the individuals of
a new generation are highlighted:

• Crossover children are created by combining pairs of parents in the cur- rent
population;

• Mutation children are created by randomly changing the genes of the indi-
vidual parents;

• Elite children are the individuals in the current generation characterized by


the best fitness values. The number of elite children changes depending on
the specific algorithm considered;
3.4.2 Mutation

Mutation is an important method of preserving the diversity of the solution candidates. In


fixed-length string chromosomes, it can be achieved by modifying the value of one element
126 3 Genetic Algorithms
of the genotype, as illustrated in Figure 3.3. More generally, a mutation may change 0 ≤ n <
len(g)
split at two points, constructing a new offspring by locations in number
using parts the string.
oneInand
binary
threecoded
fromchromosomes, for example, the elements are
1.6 Multi-Objective Optimization
the first, and the middle part from the second bits which are
ancestor. Thesimply toggled.
generalized formForofreal-encoded 13
this tech- genomes, modifying an element gi can be
done by replacing
nique is the n-point crossover, also called multi-point �it with
crossover � a number
(MPX). drawn from a normal distribution with expected value
For fixed-length
1 , like
strings, the crossover points for both parents gare always ∼ N g1 , σ 2 .
ginewidentical.

Single-Point Crossover Two-Point Crossover Change One Locus


... Change n Loci

Fig. 3.5: Crossover(a)(recombination) Fig. 3.3:


of fixed-length string Value-altering mutation of string chromosomes.
chromosomes.
Crossover children (b) Mutation children

Figure 1.3: Crossover and mutation children.


3.4.3 Permutation
3.5 Variable-Length
Two common String type ofChromosomes
crossover and mutation
The permutation reproduction
operation strategies
is an alternative are method
mutation presentedwhere the alleles of two
genes are exchanged. This, of course, makes only sense if all genes have similar data types.
Figure 1.3.
3.5.1 Creation Permutation is, for instance, useful when solving problems that involve finding an optimal
The operations outlined in the ofblock
sequence items,scheme 1.4 are salesman
like the travelling
Variable-length strings can be created by first randomly drawing a length l > 0 and then
repeated until[799,
problem a stopping
800]. Here, a genotype g could
criterion
creating a list of that is metfilled
length during
with the process;
encode
random usually the algorithm stops when the maximum
the
elements.
sequence in which the cities are visited. Exchanging two alleles then equals of
switching two cities in the route.
number of generations has been reached.
3.5.2 Mutation

1.6 Multi-Objective Optimization


If the string chromosomes are of variable length, the set of mutation operations introduced
in Section 3.4 can be extended by two additional methods. On one hand, we could insert
a couple of elements at any given position into a chromosome. One the other hand, this
operation can beIn reversed
the engineering
by deleting practice oftenthehappens
elements from that the that
string. Operations problem
change to be addressed is
the length of a genotype are the reason why variable-length strings need Exchange the alleles of two genes
to be constructed
characterized by a multi-objective nature. In those cases a single-objective problem
of elements of the same type. There is no longer a constant relation between locus and
formulation,
type. It should be noted that even with
both, several
insertion constraints,
and deletion, Fig.
are may
3.4:
also not performed
adequately
Permutation
implicitly applied to a represent the
string chromosome.
problem
by crossover. itself.two
Recombining If identical
so, the right
strings representation of the
with each other can, objective
for example, leadfunction
to is a vectorial
deletion of genes. The crossover of different strings may turn out as an insertion of new
one:
genes into an individual.
3.4.4 Crossover
F ( x ) = [ F ( x1 ), F ( x2 ), . . . , F ( xk )] (1.17)
Amongst all evolutionary algorithms, genetic algorithms have a recombination operation
The relative importance of theseisobjectives
which isnatural
closest to the not generally know3.5until
paragon. Figure thethesystem
outlines recombination of two string
chromosomes, the so-called crossover, which is performed by swapping parts of two geno-
best capabilities are determined types. and trade-offs between the objectives found. Multi-
objective optimization is concerned with the crossover,
When performing minimization of a vector
both parental of objectives
chromosomes are split at a randomly deter-
Insert Delete point. Subsequently, a new child genotype is created by appending the first
mined crossover
F(x) that can be subjected part to aofnumber of constraints or bounds, as in the case
the first parent with the second part of the second parent. This method is called
of the mono-objective
Fig. optimization
3.6: Mutation of variable-length problem
string
single-point formulated
chromosomes.
crossover (SPX9 in Eq.
). In two-point 1.1 (n
crossover being
(TPX), boththe
parental genotypes are
number of the decision variables,
9 whereasis m
This abbreviation alsothe
usednumber
for simplexof conflicting
crossover, objective
see Section 16.4.
functions) :
3.5.3 Crossover
minn F ( x ), F ∈ Rn (1.18)
x ∈R

subject string
For variable-length to thechromosomes,
constraints: the same crossover operations are available as for
fixed-length strings except that the strings are no longer necessarily split at the same loci.
• equality constraints, Gi (x) = 0 ( i = 1,. . . ,me );

• inequality constraints, Gi (x)≤ 0 (me +1,. . . ,m);

• parameter bounds like xl ≤ 0 x ≤ xu , where xl is the parameters lower bound


and xl is the parameters upper bound respectively.

Note that because F ( x ) is a vectorial function, if any of the components of F ( x )


are competing, there is no unique solution to this problem. For this reason, it is
14 1. Optimization techniques and Evolutionary algorithms
1.6 Multi-Objective Optimization

Figure 1.2: Simple genetic algorithm scheme.


Figure 1.4: Typical framework of an Evolutionary Algorithm.
The operations outlined in the block scheme 1.2 are repeated until a stop-
ping criteria are met during the process; usually the algorithm stops when the
maximum number of generations has been reached.
important to introduce some Pareto3 optimality concepts, which are introduced in
the following 1.6
paragraph.
Multi-Objective Optimization
In the engineering practice often happens that the problem to be addressed
is characterized by a multi-objective nature. In those cases a single-objective
1.6.1 Paretoproblem
concepts
formulation, even with several constraints, may not adequately represent
the problem itself. If so, the right representation of the objective function is a
Optimization problems
vectorial one: involving multiple, conflicting objectives are often
approached by aggregating the objectives into a scalar function and solving the
F (x) = [F1 (x), F2 (x), ..., Fk (x)] (1.12)
resulting single-objective optimization problem. In contrast, in this study, we are
The relative importance of these objectives is not generally know until the
concerned with finding a set of optimal trade-offs, the so-called Pareto-optimal set.
In the following, we formalize this well-known concept and also 15 define the

difference between local and global Pareto-optimal sets.


A multiobjective search space is partially ordered in the sense that two arbitrary
solutions are related to each other in two possible ways: either one dominates the
other or neither dominates.
Mathematically, a Multi Objective Optimization Problem (MOOP) minimizes
the components of a vector F(x) = (f1 (x); f2 (x);. . .; fm (x)), where x is an n-dimensional
decision variable vector x ∈ N, subject to gi (x) ≤ 0; i = 1;. . .;k. The evaluation
function F : N → M maps vectors x = (x1 ;. . . ;xn ) of the decision variable space N
to vectors y = (y1 ;. . .;ym ) of the objective function space M (Figure 1.5).
The solution to a MOOP is a (possibly uncountable) set of decision variable
vectors in N: the components of the corresponding vectors in M represent the
3 Vilfredo Pareto (1848-1923): Italian engineer, economist and sociologist.
1.6 Multi-Objective Optimization 15

F1
x1
Λ = {y ∈ℜ }
k
Ω = {x ∈ℜ }
n

F
F2

F3

x2

"Decision Variable Space" "Objective Function Space"

Figure 1.5: MOOP evaluation mapping.

best trade-offs in the objective function space. For the reader’s convenience, the
rest of the section is devoted to the mathematical definitions of the key concepts
concerning Pareto optimality; examples of these concepts can be found elsewhere
[19, 52].
Pareto Dominance: A vector u = (u1 ,. . . ,um ) is said to dominate v = (v1 . . .,vm )
(denoted by u ≺ v) if and only if ∀i 1,. . . , m, ui ≤ vi ∧ ∃ j ∈ 1, . . . , m : u j < v j . u is
also said to cover v(u  v) if and only if u ≺ v or u = v.
Pareto Optimality: A solution x ∈ N is said to be Pareto optimal with respect
to the whole set N if and only if there is no other solution x 0 ∈ N for which F(x’)
dominates F(x).
Pareto Optimal Set: For a given MOOP evaluation function F : N → M, the
Pareto optimal set (POS) is defined as the subset of all the Pareto optimal vectors
in the decision variable set:

POS := {x ∈ N :=6 ∃x0 ∈ F(x0 ) ≺ F(x)} (1.19)

Pareto Front: For a given MOOP evaluation function F : N → M and Pareto


optimal set POS, the Pareto front (PF) is defined as the set of the vectors mapped
from POS to M by F:

PF := {y = F(x) = (f1 (x), . . . , fm (x)) : x ∈ POS} (1.20)

1.6.2 Traditional Solution Methods


The classical approach to the solution of multi-objective optimizations is based
on the idea to transform the original problem in a mono-objective one, with the
16 1. Optimization techniques and Evolutionary algorithms

purpose to apply then a mono-objective optimization strategy. Several methods


are based on this concept:

• Goal method: with this strategy, the equivalent scalar function to be opti-
mized is the distance between the vectorial objective function value and a
certain reference vector (Goal vector) F Goal chosen by the designer:

min d( x ) = k F ( x ) − F Goal k p (1.21)

where
k
kvk p = ( ∑ |v| p )1/p , 1≤p≤∞ (1.22)
i =1

is the classic vector norm.

• Weights method: in this case the objective function is a linear combination


of the objectives:
k
min ∑ ωi Fi ( x ) (1.23)
i =1

where ω = (ω1 , ω2 , . . . , ωk )T ∈ Rk+ (≥ 0) is the normalized weights vector


∑ik=1 ωi = 1. The weights are chosen by the designer.

• Trade off method: one objective function Fl ( x ) is chosen and the remaining
objectives Fi ( x )(i = 1, 2, . . . , k, i 6= l ) are dealt as constraints, imposing their
upper bound values ei . The problem to be solved is the follow:

min Fl ( x ), l ∈ {1, 2, . . . , k }
(1.24)
Fi ( x ) ∈ ei , ∀i = 1, 2, . . . , k i 6= l

If the point x ∗ is the unique problem solution for at least an l ∈ 1, 2, . . . , k and


ei = Fi ( x ∗ ) for all i 6= l, x ∗ is a Pareto optimal point for the original problem.

These are only few scalarization method suggested in literature. For more details
about these and other scalarization methods see references [100], [88], [87].

1.6.3 Multi-Objective Evolutionary Algorithms


The main disadvantage of all the methods discussed in section 1.6.2 is the
introduction of arbitrariness in the scalarization process; the solution obtained
is highly sensitive to the decisions made by the designer about the scalarization
(as the weight vector for example) and demands that the user have knowledge
about the underlying problem. Moreover, in solving multi-objective problems,
designers may be interested in a set of Pareto optimal points, instead of a single
point solution. Since Evolutionary algorithms work with a population of points, it
seems natural to use them in multi-objective optimization problems to capture a
number of solutions simultaneously.
The ability of genetic algorithms to simultaneously search different regions of a
solution space makes it possible to find a diverse set of solutions for very difficult
1.6 Multi-Objective Optimization 17

problems with non-convex, discontinuous and multi-modal solutions spaces. The


crossover operator may exploit structures of good solutions with respect to different
objectives to create non-dominated solutions in unexplored parts of the Pareto
front. In addition, most multi-objective Evolutionary algorithms do not require the
user to prioritize, scale or weigh the objectives. Therefore Evolutionary algorithms
have been the most popular heuristic approach to multi-objective design and
optimization problems.
Multi-objective evolutionary algorithms (MOEA) differ from their mono-objective
version about the fitness assignment; the most effective fitness assignment strategy
for multi-objective optimization is the Pareto ranking approach which explicitly
utilize the concept of Pareto dominance (section 1.6.1) [39] in evaluating fitness
and assigning selection probability to solutions.

The population is ranked according to a dominance rule, and than each individual is
assigned a fitness value based on its rank in the population, not its actual objective
function value

For more details about multi-objective optimization using evolutionary algorithms,


see references [23], [19], [31].
18 1. Optimization techniques and Evolutionary algorithms
Chapter 2

GeDEAII: A powerful and robust


MOEA

In this chapter the Evolutionary Algorithm designed and developed during my


PhD course, namely the GeDEAII, will be presented. It shares the same framework
of its precedessor (presented in [14]), whereas it features a new crossover operator,
the Simplex-Crossover, and a novel mutation operator, the Shrink-Mutation. Finally,
a new step was added to the ones described in the GeDEA reference paper: the
Self-tuning operator. GeDEM operator was left unchanged and completed using
the non-dominated-sorting based on crowding distance. The comparison among
GeDEA-II and GeDEA, as well as with three other modern elitist methods, on
different extremely multidimensional test problems, clearly indicates that the
performance of GeDEA-II is, at least in these cases, superior.

2.1 Multiobjective Evolutionary Algorithms and the Prob-


lem of Diversity Preservation
The search process for the solutions to a MOOP has two main objectives:
1. To drive the search towards the true Pareto-optimal set/front;
2. To prevent premature convergence and distribute the solutions along the
set/front itself.
Therefore, MOEAs have to face two major issues in order to accomplish these
tasks, that is, respectively:
1. How to perform fitness assignment and selection (using aggregating, non-
Pareto population-based or Pareto-based approaches);
2. How to maintain genetic diversity within the population of solutions.
To the best of the authors’ knowledge, these requirements are nowadays fulfilled
(in different manner) by most of state-of-the-art MOEAs.
However, two additional features are demanded of an efficient and robust
MOEA, which are:

19
20 2. GeDEAII: A powerful and robust MOEA

1. To perform the optimization containing at a minimum the overall optimiza-


tion time;

2. To perform multiple runs while achieving the same results.

The latter feature is referred to as repeatability, and it is important in order


to judge the efficiency of the evolutionary algorithm under investigation. In the
following, we briefly analyze some of the most popular strategies used by the
respective MOEAs to promote diversity. For comprehensive overviews of evolu-
tionary approaches to multi-objective optimization the reader is referred to the
following more specific studies [23, 19].
Many evolutionary algorithms for multiobjective optimization have been pro-
posed [102, 23, 21]. Probably, the most popular multiobjective evolutionary algo-
rithms today are the Strength Pareto Evolutionary Approach-2 (SPEA-2) [104] and
the Nondominated Sorting Genetic Algorithm-II [24], which we have used for
comparison in the Experimental results section. The two algorithms similarly main-
tain a separate population of size N (current population, or offspring population)
and a fixed-capacity archive (previous population, or parent population), often
(as in NSGA-II) also dimensioned N. In each generation, the current population
is merged with the archive. This merged set is then subject to the process of elite
preservation, wherein the individual solutions are ranked. The new archive is
filled by taking the best-ranked solutions from the merged list. Rank conflicts are
resolved via a diversity metric. Individuals are also subject to tournament selec-
tion, crossover, and mutation to form the population for the next generation. The
main difference between the two is the way elite preservation is applied. NSGAII
invokes a procedure called nondominated sorting. Nondominated sorting assigns
domination ranks to each individual solution in a population, in such a manner
that solutions are assigned lower ranks than the ones they dominate. Solutions are
further prioritized based on a diversity metric that calculates the largest enclosing
objective space hypercube that does not contain any other solutions with the
same non-dominated sorting rank. The perimeter is then used to compute the
density around each solution. SPEA-2, on the other hand, uses a procedure for
identifying and preserving elites, first by calculating the strength of each solution
as the number of solutions it dominates, and then summing the strengths of other
solutions that dominate it as its raw fitness. A diversity term is then added to this
to obtain the final fitness.
Another interesting multiobjective evolutionary algorithm is the Indicator-
Based Evolutionary Algorithm (IBEA), proposed in [103]. It is a multiobjective
evolutionary algorithm that calculates fitness values by comparing individuals
on the basis of a quality indicator. Thereby, no particular diversity preservation
technique such as fitness sharing, clustering, etc. is necessary.
2.2 Genetic Diversity Evolutionary Algorithm
(GeDEA) 21
2.2 Genetic Diversity Evolutionary Algorithm
(GeDEA)
As GeDEA (Genetic Diversity Evolutionary Algorithm) forms the basis for
GeDEA-II, we give a brief summary of the algorithm here. For a more detailed
description, the interested reader is referred to [14]. The Genetic Diversity Evo-
lutionary Algorithm (GeDEA) is a framework that is strictly designed around
GeDEM to exalt its characteristics. Some of the design choices follow from the
basic features of GeDEM (e.g., the replacement of clones, the use of an elitist
strategy), the others are inspired by the will to make things as simple as possible,
and neither introducing arbitrary parameters nor using sophisticated heuristics.
To briefly introduce the GeDEM principle, it is worth to conceptually go back to
the section 2.1, where it has been stated that that the multi-objective optimization
process has two objectives, which are themselves conflicting: the convergence to
the Pareto-optimal set and the maintenance of genetic diversity within the pop-
ulation. The basic idea of GeDEM is to actually use these objectives during the
evaluation phase and to rank the solutions with respect to them, emphasizing the
non-dominated solutions as well as the most genetically different.

When the GeDEM is applied, the actual ranks of the solutions are determined
maximizing (i) the ranks scored with respect to the objectives of the original MOOP,
the non-dominated solutions having the highest rank, and (ii) the values assigned
to each individual as a measure of its genetic diversity, calculated according
to the chosen distance metric, i.e. the (normalized) Euclidean distance in the
decision variable space. The structure of GeDEA follows the main steps of a (µ + λ)
Evolution Strategy [10]. The evolution, however, is considered at its genotypic level,
with the traditional binary coding of the decision variables. In the following the
framework of the GeDEA is recalled for clarity.

Step 1: An initial population of µ individuals is generated at random.

Step 2: A mating pool of 2λ individuals is formed, each individual having the


same probability of being selected.

Step 3: λ offspring are generated by crossover. Some bits of the offspring are also
randomly mutated with a probability pmut .

Step 4: The whole population of µ + λ individuals is checked to discover possible


clones. These clones are removed and replaced with new randomly generated
individuals (this is done to encourage the exploration of the search space
and also to have the algorithm evaluate, for convenience, new λ different
offspring every generation; still the occurrence of clones’ birth is not so
frequent if clones are removed generation after generation).

Step 5: The objective function values of the µ + λ individuals are evaluated and
the non-dominated sorting procedure by Goldberg (1989) is performed to
assign the ranks to the solutions according to the objectives of the MOOP.
22 2. GeDEAII: A powerful and robust MOEA

Step 6: The whole population of µ + λ individuals is processed to determine the


value of the distance-based genetic diversity measure for each individual.

Step 7: GeDEM is applied according to the ranks scored in Step 5 and the values
of the diversity measure assigned in Step 6. The non-dominated sorting
procedure by [39] is used again to assign the ranks.

Step 8: The best µ solutions among parents and offspring, according to the ranks
assigned in Step 7 by GeDEM, are selected for survival and the remaining λ
are eliminated.

Step 9: If the maximum number of generations is reached then stop, else go to


Step 2.

While GeDEAII shares with its predecessor the same framework, it is presented
in this work in a real-coded fashion, hence it features the same parameters represen-
tation of the competitor algorithms already presented in Section 2.1. Moreover, this
choice was made in view of using it for solving real-world engineering optimiza-
tion problems. In order to prove the efficiency of the (µ + λ) Evolution Strategy,
whose steps are strictly followed in GeDEAII, a comparison with the competitor
algorithms framework was performed. For each algorithm, the same crossover,
mutation and selection operators were exploited, that is the SBX Crossover [2],
the Polynomial Mutation (implemented as described in [25]) and the Tournament
Selection [9] operators, respectively. Results of these comparisons are depicted in
Figure 2.1. The results presented here refer to the ZDT1 bi-objective test problem,
which involves 30 decision variables, and is thoroughly described in Section 2.4.1.
Results hint that the GeDEAII framework provides remarkable results in terms
of both convergence and Pareto Set coverage, and therefore underlies the good
performance of the algorithm itself.

2.3 Genetic Diversity Evolutionary Algorithm-II


(GeDEA-II)
GeDEA proved to be an efficient algorithm, able to explore widely the search
space, while exploiting the relationships among the solutions. In order to enhance
GeDEA algorithm performance further, several main features were added to
the previous GeDEA version, yet retaining its constitutive framework. The main
innovation is the novel crossover function, namely the Simplex-crossover, which
takes place in lieu the previous Uniform crossover. Novel selection and mutation
operators were also developed. The first one, namely the Neighbour-selection
operator, allows exploring the design space more effectively. The second one,
namely the Shrink-mutation, allows exploring more effectively the design space.
Finally, a new step was added to the ones described in the GeDEA reference
paper ([14]): the Self-tuning operator. The remaining steps characterizing GeDEA
algorithm, in particular the GeDEM, were left unchanged. The latter was integrated
with the Non-Dominating sorting procedure based on the crowding distance.
2.3 Genetic Diversity Evolutionary Algorithm-II
(GeDEA-II) 23
1
True Pareto Front
GeDEA−II
0.9
GeDEA
IBEA
0.8 NSGAII
SPEA2

0.7

0.6

0.5

f2
0.4

0.3

0.2

0.1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1


f1

Figure 2.1: Approximate Pareto-optimal set reached by GeDEAII, GeDEA, IBEA, NSGA-II
and SPEA2 provided with the same evolutionary operators, on ZDT1 test function.

2.3.1 The SIMPLEX crossover

In many Evolutionary Algorithms (EAs), a recombination with two parents is


commonly used to produce offspring. At the end of the 90’s, in several studies
the use of more than two parents for recombination in EAs have been reported
[27, 65, 90].
In 1999, Tsutsui et. al [91] proposed the simplex crossover (SPX), a new multi-
parent recombination operator for real-coded GAs. The experimental results with
test functions used in their studies showed that SPX works well on functions having
multimodality and/or epistasis with a medium number of parents: 3-parent on a
low dimensional function or 4 parents on high dimensional functions. However,
the authors did not consider the application of the SPX to multiobjective problems.
Moreover, they did not consider the possibility to take into account the fitness of
the objective function/s as the driving force of the simplex. Therefore, we decided
to integrate in the GeDEA-II the simplex crossover with these and further new
distinctive features.
Before introducing the Simplex crossover exploited in the GeDEA-II, some
words are spent to elucidate the Simplex algorithm, whose first release was pre-
sented in [84] . A simplex in n-dimensions is a construct consisting of n+1 solutions
xk , k = 1, 2, . . . , n + 1 [64]. In a two dimensional plane, this corresponds to a tri-
angle. The solutions are evaluated in each step and the worst solution w, i.e., the
one with the highest fitness value, is identified. The centroid, M , of the remaining
points is computed as M = n1 ∑ x xk (k identifying the two best solutions) and a
new solution r, replacing w, is obtained by reflection, r =M+(M−w). In the Nelder
and Mead version of the simplex algorithm, further operators are considered, such
as expansion, internal contraction and external contraction. However, they are
not taken into account in this work, since this choice would result in additional
24 2. GeDEAII: A powerful and robust MOEA

functions evaluations, as well as add complexity to the algorithm. In Figure 2.2,


the reflection step of the Nelder and Mead simplex algorithm is depicted, applied
to a problem in R2 .
w is the worst point, to be replaced by point r. M is the centroid between the
two other points, x1 and x2 .

��

��� �� ��


Figure 2.2: The reflection step of the simplex algorithm applied to a problem in R2 .

GeDEA-II utilizes the Simplex concept as the crossover operator, in order to


speed-up the evolution process, as stated below. As a matter of fact, crossover
function plays an important role in the EAs, since it combines two individuals,
or parents, to form a new individual, or child, for the next generation. Since the
Simplex is itself an optimization algorithm, the generated children are expected
to feature best fitness values when compared to the parents. Unlike the Simplex-
crossover presented in [91], Simplex-crossover exploited in GeDEA-II requires only
two parents to form a new child. This choice was motivated by the following
considerations. First of all, it is reminded here that two is the minimum number
required to form a simplex. From linear algebra, it can be easily demonstrated4
that, given k vectors, there can be found more couples of mutually linearly indepen-
dent vectors than can be done when considering triplets (or, in general, n-tuples)
of independent vectors. As a straightforward consequence, it follows that this
statement is even more true if not independence but only diversity (that is, at least
one component different from a vector to another one) is required. Therefore, every
time a new child is created, this characteristic of the Simplex-crossover ensures that
this child comprises genes different from those featuring the other children, and
allows the greatest design space exploration, due to the diversity of the parents.
4 Let us consider three vectors in R3 design space, namely ~a=(1,0,0), ~
b=(0,1,0), and ~c=(0,0,1). There
can be found three couples of linearly independent vectors, that is [~a,~b], [~a,~c] and [~b,~c] but only a
triplet of mutually and simultaneously independent vectors, that is the triplet [~a,~b,~c]. This simple
demonstration remains valid when extended to the Rn space.
2.3 Genetic Diversity Evolutionary Algorithm-II
(GeDEA-II) 25
These two parents are selected according to the selection procedure from the previ-
ous population, and combined following the guidelines of the simplex algorithm.
Let assume p1, p2 being the two parent vectors, characterized by different, multiple
fitness values, the child vector Child is formed according to the reflection move
described in ([64]:
Child := (1 + re f l ) · M − Re f l · p2 (2.1)

where Child is the new formed child and Refl is the reflection coefficient.
It is assumed that p1 is the best fitness individual among the two chosen to
form the Child, whereas p2 the worst one. Belove the strategy followed to decide
every time the best and the worst individual is highlighted.
M is the centroid of p1, calculated in the following manner:

1
 
M := · ( p1 ) (2.2)
n
where n is the number of the remaining individual, excluded the worst one. A
dedicated discussion will be done concerning this coefficient, and reported belove.
Refl coefficient is set equal to a random number (re f l ∈ [0, 1]), unlike the elemental
Simplex theory, which assumes a value equal to 1 for the Refl coefficient. This
choice allows to create a child every time distant in a random manner from the
parents, hence to explore more deeply the design space.
Moreover, unlike the Simplex algorithm theory, it was decided to give the
algorithm the possibility to switch between 1 and 2 the coefficient n in Equation 2.2.
This choice proved to be effective when used in conjunction with the Self-tuning
operator described in Section 2.3.4, since the algorithm is given the possibility to
choose every time the proper parameter setting.
The information about the fitness values is the key issue of this version of
crossover: unlike the Simplex-crossover operator presented in Tsutsui et. al [91],
this characteristic allows the crossover process to create a new individual, which is
expected to be better than the parents. This new crossover operator was expected to
combine both exploration and exploitation characteristics. In fact, the new formed
child comprises the genes of two parents, that means a good exploration of the
design space. However, it explores a design space region opposite to that covered
by the parent number 2, that means it explores a region potentially not covered
so far. In the early stages of the evolution, this means that child moves away from
regions covered from bad parents, while exploring new promising ones.
Since the Simplex algorithm is itself a single-objective optimizer, a strategy was
implemented to adapt it to a multi-objective algorithm. To deeply exploit the char-
acteristics of the simplex, at each generation the mean of each objective function,
extended to all the µ individuals, is computed. This mean is then compared to
the one characterizing the previous generation, and the objective function having
the greatest variation is selected as the fitness function used within the Simplex
algorithm to decide every time the best and the worst individual. This choice was
made after several experiments, which showed how a correct balance between
exploration of the search space and the convergence to the P.F. can be achieved by
26 2. GeDEAII: A powerful and robust MOEA

means of a switching among multiple objective functions, each time selecting the
most promising one.
In Figure 2.3, the flow chart related to the application of the Simplex crossover in
a multi-objective context is presented, extended to the most general case involving
M objective functions. It it assumed that all of the objectives are to be minimized.
At each generation ignr, the mean of each objective function mean is calculated.
Based on these values, the percentage variations PV are subsequently derived. At
this point, the two selected parents are sorted according to these values, and the
child created according to Eqs. 2.1 and 2.2. This choice guarantees that the objective
function characterized by the greatest variation is selected every time, therefore
ensuring the highest convergence rate to the PF. For test problem involving more
than two objective functions, the objective function considered to form the new
child is chosen randomly in order to enhance the design space exploration of the
crossover, required in highly dimensional objective spaces.

In order to demonstrate the effectiveness of the Simplex Crossover, two optimiza-


tion runs, regarding two test problems, were performed, where each competitor
MOEA was equipped first with the Simplex Crossover and then with the Simu-
lated Binary crossover (while keeping the other operators unchanged). Two figures
are presented that show the Approximate Pareto-optimal Sets reached from the
competitors on two test functions selected among those presented Section 2.4.1.
Both the two test problems involve 100 decision variables. In the first problem,
regarding ZDT6 , and shown in Figure2.4(a), Simplex Crossover is used instead of
Simulated Binary crossover (described for the first time in [2]), with a probability
of 75 percent. In Figure2.4(b), concerning the DTLZ6 test problem, this probability
is increased up to 90 percent. Results hint that the Simplex Crossover is very
powerful when used in conjunction with the SBX in speeding up the evolution
process without penalizing design space exploration. During evolution, GeDEA-II
makes use exclusively of the Simplex Crossover until half of the generations has
been reached. After that, Simplex Crossover is used alternatively with the SBX
with a switching probability of 50 percent. This choice is motivated by the will of
improving further the distribution and uniformity of the candidate solutions on
the Approximate Pareto-optimal Set.

2.3.2 The Shrink mutation


As far as mutation is concerned, a new Shrink-mutation operator is introduced
in the GeDEA-II.
In the open literature, this kind of mutation strategy is referred to as Gaussian
mutation [11], and conventional implementations of Evolutionary Programming
(EP) and Evolution Strategies (ES) for continuous parameter optimization using
Gaussian mutations to generate offspring are presented in [10] and [30], respec-
tively.
2.3 Genetic Diversity Evolutionary Algorithm-II
(GeDEA-II) 27

for i=1:M
q
MEANi (ignr ) =
old f it2h,i
µ
∑ h =1
µ END for

for i=1:M
MEANi (ignr −1)− MEANi (ignr )
PVi = MEANi (ignr −1)
END for

Update parents Choose two parents, p1 and p2, ac-


candidates cording to the Selection operator

Store in old f it(i, j) the fitness


value related to the ith objec-
tive function of the jth parent

A = maxi∈ M ( PVi )

Find index K ∈ M correspondent to A

Set OLDFIT = [old f it(K, 1), old f it(K, 2)]


no

Sort parents p1, p2, according


to the values stored in OLDFIT

Is the offspring
Perform Simplex crossover according to
population
the guidelines given in Eqs.2.1 and 2.2
completed?

yes

stop

Figure 2.3: Application of Simplex crossover in a multi objective context.


28 2. GeDEAII: A powerful and robust MOEA

ZDT6 test function.


Simplex Crossover(SC) vs. Simulated Binary Crossover(SBX) crossover DTLZ6 test function. DTLZ6 test function.
6 Simplex Crossover(SC) crossover Simulated Binary Crossover(SBX) crossover

IBEA_SC IBEA_SBX
SPEA2_SC 90 SPEA2_SBX
5 1 NSGA−II_SC NSGA−II_SBX
True P.F. 80 True P.F.
0.9
IBEA_SC 0.8 70
4
NSGA−II_SC
SPEA2_SC 0.7 60
IBEA_SBX 0.6
3
f2

NSGA−II_SBX 50
SPEA2_SBX 0.5

f3
f3
True P.F. 40
0.4
2 30
0.3

0.2 20
1
0.1 10

0 1 0 100
0 0 0
0.5 50
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.5 50
f1 1 0 f2 100 0
f2
f1 f1

(a) ZDT6 test function (b) DTLZ6 test function

Figure 2.4: Approximate Pareto-optimal set reached by IBEA, NSGA-II and SPEA2 pro-
vided with and without the SimplexCrossover, on ZDT6 (on the left) and DTLZ6 (on the
right) test functions.

In general, mutation operator specifies how the genetic algorithm makes small
random changes in the individuals in the population to create mutation children.
Mutation provides genetic diversity and enables the genetic algorithm to search
a broader space. Unlike the previous version of mutation featuring GeDEA algo-
rithm, where some bits of the offspring were randomly mutated with a probability
pmut , here the mutation operator adds a random number taken from a Gaus-
sian distribution with mean equal to the original value of each decision variable
characterizing the entry parent vector. The shrinking schedule employed is:
 
ignr
Shrink i := Shrink i−1 · 1 − (2.3)
ngnr

where Shrink i is a vector representing the current mutation range allowed for that
particular design variable, ignr represents the current generation and ngnr the
total number of generations. The shape of the shrinking curve was decided after
several experimental tests. The fact that the variation is zero at the last generation
is also a key feature of this mutation operator. Being conceived in this manner,
the mutation allows to deeply explore the design space during the first part of
the optimization, while exploiting the non-dominated solutions during the last
generations. Once the current variation range has been calculated, one decision
variable of the mutated child is randomly selected, and mutated according to the
following formula, resulting from an extensive experimental campaign:
p
Childmut := Childcross + [( Shrink i · random) · Shrink i ] (2.4)

where Childmut is the mutated decision variable, Childcross is the decision variable
generated by the previously introduced crossover operator and random is a ran-
dom number taken from a normal distribution in the open interval [-1,1]. Unlike
crossover operator, which generates all the offspring, mutation is applied only
2.3 Genetic Diversity Evolutionary Algorithm-II
(GeDEA-II) 29
DTLZ1 test function. DTLZ1 test function. DTLZ1 test function. DTLZ1 test function.
Shrink mutation Polynomial mutation Shrink mutation Polynomial mutation
2.5 1
5 10

0.9
4.5 9

2 IBEA 0.8
4 Spea2 8
True P.F. IBEA
0.7 True P.F.
True P.F.
3.5 7
Spea2
1.5 0.6 True P.F.
3 6

0.5

f3
2.5 5
f3

f3

f3
1 0.4
2 4
0.3
1.5 3

0.5 0.2
1 2
0.1
0.5 1

0 0
0 0 0
0
0 0
1 0.5
2 5
2 1.5 2 2.5 1
0.5 1 1 0.5 4 4 5 10
0 0 2 3 10 5
0 1 0
f1 f2 f1 f2
f1 f2 f1 f2

(a) DTLZ1 test function. IBEA algorithm (b) DTLZ1 test function. Spea2 algorithm

DTLZ1 test function. DTLZ1 test function.


Shrink mutation Polynomial mutation

5 10

4.5 9

4 8

3.5 7 Nsga2
Nsga2
3 True P.F. True P.F.
6

2.5 5
f3

f3

2 4

1.5 3

1 2

0.5 1

0 0
0 0
2 5
4 10
4 2 5
0 10 0
f1 f2 f1 f2

(c) DTLZ1 test function. Nsga2 algorithm

Figure 2.5: Approximate Pareto-optimal sets reached by IBEA, NSGA-II and SPEA2
provided with the ShrinkMutation operator (on the left side of each subfigure) and with
the PolynomialMutation operator (on the right side of each subfigure).

on a selected part of the offspring, depending on the strategy selected by means


of the Self-tuning operator. Before starting offspring mutation, offspring popula-
tion is randomly shuffled to prevent locality effects. After that, a pre-established
percentage (based on the Self-tuning operator) of the individuals are selected of
mutation. The initial Shrink factor was set equal to ten percent of the variation
range of the design variables. This mutation operator was found to be powerful
especially in multi-objective problems requiring a huge exploration of the design
space. In Figure 2.5, a comparison between Shrink mutation and Simulated Binary
mutation is depicted. Results refer to DTLZ1 test function, a tri-objective test func-
tion involving 7 decision variables, and described in Section 2.4.1. The algorithms
chosen for the test are the same introduced in Section 2.4. Similar behaviors were
found with the other test function, not reported here for brevity and clarity.
30 2. GeDEAII: A powerful and robust MOEA

Figure 2.5(a) shows that Polynomial mutation used in IBEA algorithm gives the
best results in terms of convergence towards the True Pareto Front. However, as far
as distribution of solutions is concerned, Shrink mutation performs better. However,
the results of this comparison are slightly different when NSGA-II, whose test is
presented in Figure 2.5(c) and SPEA2, whose test is presented in Figure 2.5(b),
are provided with the Shrink mutation operator. In these cases, the latter gives
the best results in terms of both convergence and distribution when compared to
Polynomial mutation. Thank to these results, as far as GeDEA-II is concerned, it
was decided to use them alternatively during the evolution, in order to capture the
best of the two mutation operators.

2.3.3 Diversity preservation


As underlined in Sections 2.1 and 2.2, maintaining the genetic diversity within
the population is mandatory for a robust EA. To this purpose, in GeDEA-II two
diversity preservation mechanism are used, namely the GeDEM, already employed
in GeDEA [14] and the Non-Dominated Sorting [24]. Both of the two mentioned
mechanism are adopted since in authors’ opinion each of them has unique features
which can take benefit from each other. To make this assertion clearer, it is worth
to briefly go back to the mathematical definition of GeDEM and non-dominated
sorting based on crowding distance.The definition of dominance used in the
non-dominated sorting procedure performed by GeDEM is:

Vector u = (ranku ; distu ) dominates vector v = (rankv ; distv )


if and only if (ranku > rankv ) (distu ≥ distv )
W

On the contrary, the definition of dominance used in the non-dominated sorting


based on crowding distance is:

Vector u = (ranku ; distu ) dominates vector v = (rankv ; distv )


if and only if (ranku > rankv ) [(ranku = rankv ) (distu ≥ distv )]
V W

Clearly, the logical operator is the great difference between the aforementioned
diversity mechanisms, which entails the slightly different behavior of the two
algorithms. In particular, GeDEM tends to create less non-dominated individuals,
since both the rank and the diversity conditions are to be fulfilled simultaneously.
Therefore, the evolution process results faster. On the other hand, non-dominated
sorting based on crowding distance tends to create more non-dominated individu-
als, which results in a better Pareto front coverage. In order to take advantage of
both the characteristics, in GeDEA-II the diversity preservation is accomplished
by means of GeDEM, in the first three quarters of the generations, whereas in the
remainder of the generations the Non-Dominated Sorting mechanism is exploited.

2.3.4 The Self-tuning operator


The last enhancement introduced in the GeDEA-II is the Self-tuning operator,
which allows to automatically switch between two set of parameters, and to give
2.3 Genetic Diversity Evolutionary Algorithm-II
(GeDEA-II) 31
the algorithm the possibility to better adapt itself to the particular multi-objective
problem being optimized.
These parameters were selected on the basis of a preliminary experimental
procedure, aiming at pointing out the most influent parameters on both exploration
and convergence characteristics of the GeDEAII.
Initially, they were organized into different sets, and after several experimental
tests two sets were identified, namely Set0 and Set1 . Therefore, the GeDEA-II
behaviour can be changed by simply switching between Set0 and Set1 .
The first parameter regards the n value in the Simplex-Crossover, 0 meaning a n
value equal to 2, 1 meaning a n value equal to 1.
The second regards the switching among the objective functions in order to
select the one which will act as the driving force of the Simplex Crossover: 0 means
that the choice comes down to the one having the greatest variation, as anticipated
in Section 2.3.1, whereas 1 means that the switching is performed in a random
manner.
The third parameter regards mutation operator, 0 referring to the original
mutation operator employed in GeDEA version (all the individuals undergo
mutation), whereas 1 referring to the new Shrink-mutation operator (10% of the
individuals undergo mutation).
Only two distinctive sets were established to make the procedure as much
simple as possible and at the same time to reduce at a minimum the number of
generations required for the optimization.
Details of the Self-tuning operator constitutive framework are given below:

1. The first four generations are calculated by means of Set1 , starting from the
initial population randomly created;

2. The second four generations are calculated by means of Set0 , once again
starting from the same the initial population;

3. The mean of all the objective functions are calculated for the initial population;
calculations are repeated for the final populations created with each set (the
fourth and the eighth population, respectively);

4. For each of the objective functions, the percentage variations are calculated,
for both the two sets.

5. For both the two sets, the greatest variation is memorized;

6. Finally, the set is selected, which yielded the greatest variation between the
two stored;

In particular, an experimental campaign showed us that four generations per


each set were sufficient to select the best set. The remaining generations are
calculated with the finally selected set.
In Table 2.1, the two sets are introduced. Despite its simplicity, this operator
proved to give the GeDEA-II the ability to well adapt itself to multi-objective
problems of different nature.
32 2. GeDEAII: A powerful and robust MOEA

Operator Set0 Set1


Crossover n=2 n=1
Crossover Best objective function Random switching
Mutation Shrink-mutation (10% of the individuals) Shrink-mutation (100% of the individuals)

Table 2.1: Settings of the Self-tuning operator.

2.4 Comparison with Other Multiobjective Evolutionary Al-


gorithms
In order to judge the performance of the GeDEA-II, a comparison with other
different state-of-the-art multi-objective EAs was performed. SPEA-2 [102], NSGA-
II [24] and IBEA [103] were chosen as competitors, and their performance against
GeDEA-II were measured on five test problems featuring the characteristics that
may cause difficulties in converging to the Pareto-optimal front and in maintaining
diversity within the population [22]: convexity, non-convexity, discrete Pareto
fronts, multimodality, and biased search spaces. In addition, their performance
were tested also on Kursawe test Function KUR [54]. The latter consists of a multi-
modal function and function with pair-wise interactions among the variables, the
Pareto front is disconnected consisting of concave and convex parts and an isolated
point. Finally, GeDEA-II performance was tested on three more recent and more
challenging benchmark test functions, i.e. the scalable Test Problems presented in
[26]. The simplicity of construction, scalability to any number of decision variables
and objectives, knowledge of the shape and the location of the resulting Pareto-
optimal front, and introduction of controlled difficulties in both converging to the
true Pareto-optimal front and maintaining a widely distributed set of solutions are
the main features of the suggested test problems. The thirteen test functions, the
methodology and the metric of performance used in the comparison are briefly
recalled in the following for easy reference.

2.4.1 Test functions


Each of the five test functions T1 , T2 , T3 , T4 and T6 introduced by [102] is a
two-objective minimization problem that involves a distinct feature among those
identified by [22]. All the test functions are constructed in the same way, according
to the guidelines in [22]:

Minimize : T (x) = ( f 1 ( x1 ), f 2 ( x ))
subject to : f 2 (x) = g( x2 ; . . . ; x m )h( f 1 ( x1 ), g( x2 ; . . . ; x m )) (2.5)
where : x = ( x1 , . . . , xM )

Function f controls vector representation uniformity along the Pareto Approxi-


mation Set. Function g controls the resulting MOP characteristics (whether it is
multifrontal or has an isolated optimum). Function h controls the resulting Pareto
front characteristics (e.g., convex, disconnected, etc.) These functions respectively
2.4 Comparison with Other Multiobjective Evolutionary Algorithms 33

influence search along and towards the true Pareto front, and the shape of a Pareto
front in R2 . Deb [22] implies that a MOEA has difficulty finding PFtrue because it
gets “trapped” in the local optimum, namely PFlocal . Test functions reported
in this work feature an increased number of decision variables, when compared to
their original versions reported in [102]. This choice was motivated by the authors’
will of testing exploration capabilities of the algorithms also on highly dimensional
test problems, and contributes to justify the results presented in Section 2.4.4.

• Test function T1 has a convex Pareto-optimal front:

f 1 ( x 1 ) = ( x1 )
n
xi
g( x2 ; . . . ; x n ) = 1 + 9 · ∑ (2.6)
i =2
( n − 1)
s
f1
h ( f1 , g ) = 1 −
g

where n = 100 and xi ∈ [0,1]. The Pareto-optimal front corresponds to g(x) =


1. The original version presented in [102] featured 30 decision variables.

• Test function T2 has a non-convex Pareto-optimal front:

f 1 ( x 1 ) = ( x1 )
n
xi
g( x2 ; . . . ; x n ) = 1 + 9 · ∑ (2.7)
i =2
( n − 1)
 2
f1
h ( f1 , g ) = 1 −
g

where n = 100 and xi ∈ [0,1]. The Pareto-optimal front corresponds to g(x) =


1. The original version presented in [102] featured 30 decision variables.

• Test function T3 features a Pareto-optimal front disconnected, consisting of


several noncontiguous convex parts:

f 1 ( x 1 ) = ( x1 )
n
xi
g( x2 ; . . . ; x n ) = 1 + 9 · ∑ (2.8)
i =2
( n − 1)
s !  
f1 f
h(f1 , g) = 1 − − 1 · sin(10πf1 )
g g

where n = 100 and xi ∈ [0,1]. The Pareto-optimal front corresponds to g(x) =


1. The original version presented in [102] featured 30 decision variables.

• Test function T4 contains 219 local Pareto-optimal fronts and, therefore, tests
34 2. GeDEAII: A powerful and robust MOEA

for the EA ability to deal with multifrontality:

f 1 ( x 1 ) = ( x1 )
n
g( x2 ; . . . ; x n ) = 1 + 10(n − 1) · ∑ xi 2 − 10 cos(4πxi )) (2.9)
i =2
s
f1
h ( f1 , g ) = 1 −
g

where n = 100 and xi ∈ [0,1]. The Pareto-optimal front is convex and cor-
responds to g(x) = 1. The original version presented in [102] featured 10
decision variables.

• Test function T6 features two difficulties caused by the nonuniformity of the


search space: first, the Pareto optimal solutions are nonuniformly distributed
along the PFtrue (the front is biased for solutions for which f 1 ( x1 ) is near
one); and second, the density of the solutions is lowest near the PFtrue and
highest away from the front::

f 1 ( x1 ) = 1 − exp(−4x1 ) sin6 (6πx1 )


!1/4
n
xi
g( x2 ; . . . ; x n ) = 1 + 9 · ∑ (2.10)
i =2
( n − 1)
 2
f1
h(f1 , g) = 1 −
g

where n = 100 and xi ∈ [0,1]. The Pareto-optimal front is non-convex and


corresponds to g(x) = 1. The original version presented in [102] featured 10
decision variables.

The multi-objective test function designed by Kursawe [54] is included because


this two-objective function PFtrue has several disconnected and unsymmetric areas
in solution space. Its PFtrue consists of three disconnected Pareto curves. Its solution
mapping into dominated objective space is quite convoluted. Its number of decision
variables is arbitrary. However, changing the number of decision variables appears
to slightly change PFtrue shape and does change its location in objective space. We
use it here with three decision variables. The variation range is augmented for all
of the variables involved by a factor two, when compared to the original range,
which is [-5,5], in order to increase the difficulty of the problem.
n −1   q 
f 1 (x) = ∑ −10 · exp −0.2 xi2 + xi2+1 ) (2.11)
i =1
n 
| xi |0.8 + 5 sin( xi3 )

f 2 (x) = ∑
i =1

Finally, the entire set of tri-objective test functions designed by Kalyanmoy Deb,
Lothar Thiele, Marco Laumanns and Eckart Zitzler, and presented in [26], are
2.4 Comparison with Other Multiobjective Evolutionary Algorithms 35

considered, in order to demonstrate the GeDEA-II capabilities on more than two-


objectives test problems. In the following, n identifies the number of decision
variables, M the number of objective functions, and k = | x M | = n − M + 1 the
number of variables of the functional g(xM ). The number of variables was always
increased when compared to that suggested by the authors in [26], whereas
the decision variables range was left unchanged. These features help clarify the
different results between those reported in Section 2.4.4 and the original ones [26].

• Test function DTLZ1 contains (11k − 1) local Pareto-optimal fronts, each of


which can attract an MOEA [26]. To increase the difficulty of the problem, the
number of variables and the frequency of the cosine function were doubled
when compared to those suggested in [26].

f 1 ( x ) = (1 + g(xM ))(x1 )(x2 )


1
f2 (x) = (1 + g(xM ))(x1 )(1 − x2 )
2
1
f3 (x) = (1 + g(xM ))(1 − x1 ) (2.12)
2 " #
and g = 100 k + ∑ ( xi − 0.5)2 − cos(40π ( xi − 0.5))
x1 ∈ x M

where n = 14 and xi ∈ [0,1].

• Test function DTLZ2 is the Generic sphere problem, according to the defini-
tion given to it in [26].

f 1 ( x ) = (1 + g(xM )) cos(x1 ß/2) cos(x2 ß/2)


f 2 ( x ) = (1 + g(xM )) cos(x1 ß/2) sin(x2 ß/2)
f 3 ( x ) = (1 + g(xM )) sin(x1 ß/2) (2.13)
2
and g = ∑ ( xi − 0.5)
x1 ∈ x M

where n = 22 and xi ∈ [0,1]. The number of variables suggested in [26] is 12.

• Test function DTLZ3 is similar to test function DTLZ2, except for the function
g, which introduces (3k − 1) local Pareto-optimal fronts, and only one global
Pareto-optimal front.

f 1 ( x ) = (1 + g(xM )) cos(x1 ß/2) cos(x2 ß/2)


f 2 ( x ) = (1 + g(xM )) cos(x1 ß/2) sin(x2 ß/2)
f 3 ( x ) = (1 + g(xM )) sin(x1 ß/2) (2.14)
" #
and g = 100 k + ∑ ( xi − 0.5)2 − cos(20π ( xi − 0.5))
x1 ∈ x M

where n = 22 and xi ∈ [0,1]. The number of variables suggested in [26] is 12.


36 2. GeDEAII: A powerful and robust MOEA

• Test function DTLZ4 is a modified version of DTLZ2, since it features a


different meta-variable mapping.
ff ff
f 1 ( x ) = (1 + g(xM )) cos(x1 ß/2) cos(x2 ß/2)
ff ff
f 2 ( x ) = (1 + g(xM )) cos(x1 ß/2) sin(x2 ß/2)
ff
f 3 ( x ) = (1 + g(xM )) sin(x1 ß/2) (2.15)
" #
and g = 100 k + ∑ ( xi − 0.5)2 − cos(20π ( xi − 0.5))
x1 ∈ x M

where n =22, α = 100 and xi ∈ [0,1]. Deb, Thiele, Laumanss and Zitzler [26]
also suggest n=12 here.

• Test function DTLZ5 features a different mapping compared to the one of


DTLZ4. This problem will test a MOEA ability to converge to a curve and
will also allow an easier way to visually demonstrate the performance of the
algorithm.

f 1 ( x ) = (1 + g(xM )) cos(x1 ß/2) cos(`2 )


f 2 ( x ) = (1 + g(xM )) cos(x1 ß/2) sin(`2 )
f 3 ( x ) = (1 + g(xM )) sin(x1 ß/2) (2.16)
" #
g = 100 k + ∑ ( xi − 0.5)2
x1 ∈ x M
π
and θ2 = (1 + 2gx2 )
(4(1 + g))

where n =22 and xi ∈ [0,1]. The number of variables suggested in [26] for this
test problem is 12.

• Test function DTLZ6 is a modified, harder-to-optimize version of the above


test problem. The number of decision variables was dramatically increased
when compared to the original one.

f 1 ( x ) = (1 + g(xM )) cos(x1 ) cos(`2 )


f 2 ( x ) = (1 + g(xM )) cos(x1 ) sin(`2 )
f 3 ( x ) = (1 + g(xM )) sin(x1 ß/2) (2.17)

g = ∑ xi )0.1
x1 ∈ x M
π
and θ2 = (1 + 2gx2 )
(4(1 + g))

where n = 100 and xi ∈ [0,1]. The number of variables suggested in [26] for
this test problem is 12.

• Test function DTLZ7 features 2 M−1 disconnected local Pareto-optimal regions


in the search space. It is chosen to test the MOEA ability in finding and
2.4 Comparison with Other Multiobjective Evolutionary Algorithms 37

maintain stable and distributed subpopulations in all four disconnected


global Pareto-optimal regions.

f 1 ( x ) = x1
f 2 ( x ) = x2
f 3 ( x ) = (1 + g(xM ))h (2.18)
9
k x1∑
g = 1+ ( xi )
∈x M
M −1  
fi
and h = M − ∑ (sin((1 + 3π f i ))
i =1
1+g

where n = 100 and xi ∈ [0,1]. Once again, the number of decision variables
was dramatically increased when compared to the original one, suggested in
[26] for this test problem, and equal to 22.

2.4.2 Methodology
The methodology used in [102] is strictly followed. GeDEA and competitors are
executed 30 times on each test function. There are different parameters associated
with the various algorithms, some common to all and some specific to a particular
one. In order to make a fair comparison among all the algorithms, most of these
constants are kept the same. In GeDEA-II, GeDEA and in competitors algorithms,
the population size is set to 100. In the following, the parameters of the competitors
MOEA are reported following the terminology used in PISA implementation5 . The
individual mutation probability is always 1 and the variable mutation probability
is fixed at 1/n, n being the number of the decision variables of the test problem
considered. The individual recombination probability along with the variable
recombination probability are set to 1. The variable swap probability is set to 0.5.
ηmutation is always set to 20 and ηrecombination is fixed to 15. For IBEA algorithm,
tournament size is always set to 2, whereas additive epsilon is chosen as the
indicator. Scaling factor kappa is set to 0.05, and rho factor is fixed to 1.1. For both
NSGA-II and SPEA2, tournament size is given a value equal to 2. NSGA-II, SPEA-2
and IBEA are run with the PISA6 implementation [15], with exactly the same
parameters and variation operators. The maximum number of generations for test
functions T1 , T2 and T6 is set to 20 for all the algorithms, for test functions T3 and
T4 the individuals are evolved for 40 generations but for the Kursawe test function
the number of generations is set to 50. For test function DTLZ1 and DTLZ7 , the
number of generations is set to 100. This number of generation is increased up
5 Individualmutation probability (probability that a certain individual undergoes mutation); individ-
ual recombination probability (probability that a certain pair of individuals undergoes recombination);
variable mutation probability (probability that a certain variable in a given individual is mutated);
variable swap probability (probability that a certain pair of variables is swapped during recombination);
variable recombination probability (probability that the SBX recombination operator is used for a given
pair of variables; this decision is independent from variable swap probability); ηmutation (distribution
index for mutation operator); ηrecombination (distribution index for recombination operator).
6 This software is available for public use at PISA website http://www.tik.ee.ethz.ch/pisa/
38 2. GeDEAII: A powerful and robust MOEA

to 150 for DTLZ3 test function. In test functions DTLZ4 and DTLZ5 , individuals
are evolved for 80 generations, whereas on DTLZ2 and DTLZ6 the number of
generation is set to 70 and 50, respectively.
In Table 2.2, the original number of generations characterizing test problems
presented in [102] and [26], is compared to the ones used here. The number
of generations was reduced in order to test the convergence properties of the
investigated algorithms. In fact, being the number of individuals constituting a
generation left unchanged, a reduction of generations directly translates into a
reduction of the objective functions evaluations.

Original version Proposed test


problems problems
ZDT1 250 20
ZDT2 250 20
ZDT3 250 30
ZDT4 250 40
ZDT6 250 20
DTLZ1 300 100
DTLZ2 300 70
DTLZ3 500 150
DTLZ4 200 80
DTLZ5 200 80
DTLZ6 500 50
DTLZ7 200 100

Table 2.2: Original and proposed number of generations for the ZDT and DTLZ test suites.

The different settings contribute to justify the different results reported here,
when compared to those presented in the original papers [102, 26].

2.4.3 Metric of Performance


Different metrics can be defined to compare the performance of EAs with
respect to the different goals of optimization itself [102]: how far is the resulting
non-dominated set from the Pareto front, how uniform is the distribution of the
solutions along the Pareto Approximation set/front, how wide is the Pareto Ap-
proximation set/front. For measuring the quality of the results, we have employed
the Hypervolume approach, due to its construction simplicity and for the reason,
which will be soon explained. The hypervolume approach by [102] (modified in
[104]) measures how much of the objective space is dominated by a given non-
dominated set. Zitzler et al. state it as the most appropriate scalar indicator since it
combines both the distance of solutions (towards some utopian trade-off surface)
and the spread of solutions. To better understand the reason for this choice, it is
worth to see at Figure 2.6 (a) and (b). The reference point is indicated as R.P..
As regards the convergence of the known P.F. to the True P.F., please consider the
case depicted in Figure 2.6 (a). The non-dominated set A has a great Hypervolume
indicator when compared to set B, due to its superior proximity to the True P.F..
As far as the spread of the solution on the Pareto Approximation Set is concerned,
2001))  measures  how  much  of  the  objective  space  is  dominated  by  a  given 
nondominated set A. Zitzler et al. state it as the most appropriate scalar indicator since 
it combines both the distance of solutions (towards some utopian trade‐off surface) and 
2.4 Comparison with Other Multiobjective Evolutionary Algorithms
the spread of solutions.  39
To better understand the reason for this choice, it is worth to see at Fig. 3 (a) and (b). 
The reference point is indicated as R.P. in Fig. 3. 
 

             

 
Fig. 3 ‐ Significance of the Hypervolume indicator as far as convergence (a, AT THE TOP), and 
Figure 2.6: Significance
 
of the Hypervolume indicator as far as convergence (a, AT THE
diversity (b,  AT THE BOTTOM) is concerned. 

TOP), and diversityAs regards the convergence of the known P.F. 


(b, AT THE BOTTOM) is concerned. to the True P.F., please consider the 
case depicted in Fig. 3 (a). The non‐dominated set A has a great Hypervolume indicator 
when compared to set B, due to its superior proximity to the True P.F.. 
As  far  as  the  spread  of  the  solution  on  the  Pareto  Front  is  concerned,  Fig.  3  (b) 
Figure 2.6 (b) qualitatively shows that a more uniform distribution of the solutions
qualitatively  shows  that  a  uniform  distribution  of  the  solutions  (on  the  right)  yields  a 
great Hypervolume indicator. 
(on the right) yields a greater Hypervolume indicator. Therefore, this indicator
Therefore,  this  indicator  is  intrinsically  able  to  compare  performances  of  different 
EAs as far as both the convergence and the genetic diversity preservation is concerned. 
is intrinsically able to compare performance of different EAs as regards both the
The  Hypervolume  Pareto  compliant  indicator  is  defined  as  the  area  of  coverage  of 
convergence to the Pareto Approximation Set and its coverage. In general, it is not
PFknown with respect to the objective space for a two‐objective MOP. This equates to the 
summation  of  all  the  rectangular  areas,  bounded  by  some  reference  point  and  (f1(x), 
sufficient for a set of candidate solutions to be closer than another one to the True
f2(x)).  Mathematically,  this  is  described  in  equation  (3)  (for  a  generic  n‐objectives 
problem): 
Pareto front, to have a higher hypervolume value. It is the blend of convergence
⎡ ⎤
                                  Hypervolume = ⎢U vol i | veci ∈ Pknown ⎥                                              (3) 
and uniformity of the final Approximation Set that counts. ⎣i ⎦
 
7
The Hypervolume is defined as the area of coverage of PF with respect
where veci is a nondominated vector in PFknown and voli is the Hypervolume (an area  known
in two objectives problems) between the reference point and vector veci. 
to the objective space for a two-objective MOP. As illustrated in Figure 2.6, this
In (Zitzler et al., 1999a; Zitzler et al., 1999b), the reference point is set to (0, 0). 
region consists of an orthogonal polytope, and may be seen as the union of
n axis-aligned hyper-rectangles with one common vertex (the reference point,
R.P.). Mathematically, this is described in equation 2.19 (for a generic n-objectives
problem):
" #
Hypervolume := (2.19)
[
voli |veci ∈ Pknown
i

where veci is a nondominated vector in PFknown and voli is the Hypervolume


(an area in two objectives problems) between the reference point and vector veci .
In this work we use the version implemented by Fonseca et al. and presented
in [32].

2.4.4 Results of Comparison


As in Zitzler et al. [102], Figures 2.7-2.12 and 2.14-2.20 show an excerpt of
the non-dominated fronts obtained by the EAs and the Pareto-optimal fronts
(continuous curves). The points plotted are the non-dominated solutions extracted
7 The Hypervolume is a Pareto compliant indicator as stated in [19].
40 2. GeDEAII: A powerful and robust MOEA

from the union set of the outcomes of the first five runs, the best and the worst
one being discarded. The performance of GeDEA-II is also compared to NSGA-II,
SPEA2 and IBEA according to the hypervolume metric defined in Equation 2.19.
The distribution of these values is shown using box plots in Figure 2.13 and 2.21.
On each box, the central line represents the median, the edges of the box are the
25th and 75th percentiles, the whiskers extend to the most extreme data points not
considered outliers, and outliers are plotted individually, with a Plus sign. Results
are normalized with the best Hypervolume value coming from the union set of all
of the runs, extended to all of the algorithms. For each test problem, the reference
point is assumed equal for all of the algorithms, and equal to the maximum value
for each objective function from the union of all of the output points. In presenting
the DTLZ test suit results, the exact order is not strictly followed due to layout
reasons.
2.4 Comparison with Other Multiobjective Evolutionary Algorithms 41
ZDT1 test function
5
GeDEA-II
4.5 GeDEA
IBEA
4
NSGA-II
3.5 SPEA2
True P.F.
3
f2

2.5

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
f1
Figure 2.7: Test function T1 (convex).

T2 test function
4.5

3.5

2.5
GeDEA−II
f2

2 GeDEA
IBEA
1.5 NSGA−II
SPEA2
1 True P.F.

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
f1

Figure 2.8: Test function T2 (non-convex).


42 2. GeDEAII: A powerful and robust MOEA

ZDT3 test function


5
GeDEA-II
GeDEA
4 IBEA
NSGA-II
SPEA2
3 True P.F.
f2

-1
0 0.2 0.4 0.6 0.8 1
f1
Figure 2.9: Test function T3 (discrete).

ZDT4 test function T4 test function


450 1
GeDEA-II GeDEA−II
400 GeDEA 0.9 GeDEA
IBEA
350 IBEA NSGA−II
0.8
NSGA-II SPEA2
300 SPEA2 True P.F.
0.7
True P.F.
250
0.6
200
f2

0.5
f2

150
0.4
100
0.3
50
0.2
0

-50 0.1

-100 0
0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
f1
f1

(a) ZDT4 test function. A proper scale in order for all (b) ZDT4 test function. True Pareto Front
the populations to be included in the plot. region.

Figure 2.10: Test function T4 (multi-modal).


2.4 Comparison with Other Multiobjective Evolutionary Algorithms 43

T6 test function
9

GeDEA−II
6 GeDEA
IBEA
5 NSGA−II
SPEA2
f2

True P.F.
4

0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
f1

Figure 2.11: Test function T6 (non-uniform).

Kursawe test function


2
GeDEA-II
GeDEA
0 IBEA
NSGA-II
-2 SPEA2

-4
f2

-6

-8

-10

-12
-20 -19 -18 -17 -16 -15 -14 -13 -12
f1
Figure 2.12: Test function KUR.
44 2. GeDEAII: A powerful and robust MOEA

ZDT1 test function T2 test function

1 1

0.9
0.9
0.8
Hypervolume-Indicator

Hypervolume−Indicator
0.7
0.8
0.6

0.7 0.5

0.4
0.6
0.3

0.5 0.2

0.1
0.4
GeDEA-II GeDEA SPEA2 NSGA-II IBEA GeDEA−II GeDEA SPEA2 NSGA−II IBEA

(a) ZDT1 boxplot (b) ZDT2 boxplot

ZDT3 test function T4 test function

1 1

0.9
0.9
Hypervolume-Indicator

Hypervolume−Indicator

0.8
0.8

0.7
0.7
0.6

0.6
0.5

0.5
0.4

GeDEA-II GeDEA SPEA2 NSGA-II IBEA GeDEA−II GeDEA SPEA2 NSGA−II IBEA

(c) ZDT3 boxplot (d) ZDT4 boxplot

ZDT6 test function Kursawe test function

1 1

0.9 0.98

0.8
0.96
Hypervolume-Indicator
Hypervolume-Indicator

0.7
0.94
0.6
0.92
0.5
0.9
0.4
0.88
0.3

0.2 0.86

0.1 0.84
GeDEA-II GeDEA SPEA2 NSGA-II IBEA GeDEA-II GeDEA SPEA2 NSGA-II IBEA

(e) ZDT6 boxplot (f) KUR boxplot

Figure 2.13: Box plots based on the Hypervolume metric. Each square contains six box plots
representing the distribution of Hypervolume values for the six algorithms. Results refer to
the ZDT test suite.
2.4 Comparison with Other Multiobjective Evolutionary Algorithms 45

DTLZ1 test function

GeDEA-II
800 GeDEA
IBEA
700 NSGA-II
SPEA2
600 True P.F.
500

f3
400

300

200

100
1000
0
0 200 500
400 600 0
800
f2
f1

(a) DTLZ1 test function. A proper scale in order for all the
populations to be included in the plot.
DTLZ1 test function

GeDEA-II
100 GeDEA
IBEA
NSGA-II
80 SPEA2
True P.F.

60
f3

40

20

100
0
0 20 50
40 60 80 100 0
f2
f1

(b) DTLZ1 test function. A proper scale in order for the


individuals lying in a region closer to the True Pareto
Front to be included in the plot.

(c) DTLZ1 test function. True Pareto Front region

Figure 2.14: Test function DTLZ1.


46 2. GeDEAII: A powerful and robust MOEA

Figure 2.15: Test function DTLZ2 .

Figure 2.16: Test function DTLZ4 .


2.4 Comparison with Other Multiobjective Evolutionary Algorithms 47

DTLZ3 test function

GeDEA-II
1800 GeDEA
IBEA
1600 NSGA-II
1400 SPEA2
True P.F.
1200

1000

f3
800

600

400

200
2000
0
0 1000
500 1000 1500 2000 0
f2
f1

(a) DTLZ3 test function. A proper scale in order for all the
populations to be included in the plot.
DTLZ3 test function

GeDEA-II
100 GeDEA
IBEA
NSGA-II
80 SPEA2
True P.F.

60
f3

40

20

100
0
0 20 50
40 60 80 0
100
f2
f1

(b) DTLZ3 test function. A proper scale in order for the


individuals lying in a region closer to the True Pareto
Front to be included in the plot.

(c) DTLZ3 test function. True Pareto Front region.

Figure 2.17: Test function DTLZ3.


48 2. GeDEAII: A powerful and robust MOEA

DTLZ5 test function

GeDEA-II
2.5 GeDEA
IBEA
NSGA-II
2 SPEA2
True P.F.

1.5
f3

0.5

0 1.5
0 1
0.5 0.5
1 0
1.5
f2
f1

Figure 2.18: Test function DTLZ5 .

Figure 2.19: Test function DTLZ7.


2.4 Comparison with Other Multiobjective Evolutionary Algorithms 49

DTLZ6 test function DTLZ6 test function

GeDEA-II GeDEA-II
GeDEA 1.5 GeDEA
100
IBEA IBEA
NSGA-II NSGA-II
80 SPEA2 SPEA2
True P.F. True P.F.
1
60
f3
f3

40
0.5

20

100 1.5
0 50 0 1
0 20 40 0 0.5 0.5
60 80 100 0 1 1.5 0
f2 f2
f1 f1

(a) DTLZ6 test function. A proper scale in (b) DTLZ6 test function. True Pareto Front region.
order for all the populations to be included
in the plot.

Figure 2.20: Test function DTLZ6.


50 2. GeDEAII: A powerful and robust MOEA

DTLZ1 test function DTLZ2 test function

1 1

0.9995 0.99

0.999
Hypervolume-Indicator

Hypervolume-Indicator
0.98

0.9985
0.97

0.998
0.96
0.9975
0.95
0.997

0.94
0.9965

GeDEA-II GeDEA SPEA2 NSGA-II IBEA GeDEA-II GeDEA SPEA2 NSGA-II IBEA

(a) DTLZ1 boxplot (b) DTLZ2 boxplot

DTLZ3 test function DTLZ4 test function

1
1

0.99
0.95
0.98
Hypervolume-Indicator

Hypervolume-Indicator
0.9
0.97

0.85
0.96

0.95 0.8

0.94
0.75

0.93
0.7
GeDEA-II GeDEA SPEA2 NSGA-II IBEA GeDEA-II GeDEA SPEA2 NSGA-II IBEA

(c) DTLZ3 boxplot (d) DTLZ4 boxplot

In general, the experimental results show that GeDEA-II is able to converge


towards the True Pareto-optimal front and to develop a widely and well dis-
tributed non-dominated set of solutions. The comparison with the other three
best-performing MOEAs according to the Hypervolume metric proves that the per-
formance of GeDEA-II is somewhat superior. Considering the specific features
of the six test functions, GeDEA-II shows similar performance both on convex
and non-convex Pareto-optimal fronts. NSGA-II, SPEA-2 and IBEA seem instead
to have more difficulties with non-convexity, since the non-dominated solutions
found on test function T1 are nearest to the Pareto-optimal front than those found
on T2 . Also discreteness (test function T3 ) does not affect the performance of
the GeDEA-II, which reach the True Pareto Front and completely covers it. The
performance of GeDEA-II is particularly remarkable in the case of multi-modality,
which is considered by Ziztler et al. [102] as the hardest problem feature. GeDEA-II
reaches the Pareto-optimal front even of test function T4 , unlike the competitors,
that are far away from the True Pareto front. GeDEA-II outperforms NSGAII and
SPEA2 in the case of biased search space (test function T6 ) and is also able to evolve
a well-distributed non-dominated set. These results gain even more significance,
since the number of decision variables was set to 100, unlike the original values of
30 (10 for the test functions T4 and T6 ).
GeDEA-II performs better than the competitors also on Kursawe test function,
even if boxplots show that its performance are slightly inferior to those it features
2.4 Comparison with Other Multiobjective Evolutionary Algorithms 51

DTLZ5 test function DTLZ6 test function

1 1

0.995
0.9
0.99

Hypervolume-Indicator
Hypervolume-Indicator

0.985 0.8

0.98
0.7
0.975

0.6
0.97

0.965
0.5
0.96

GeDEA-II GeDEA SPEA2 NSGA-II IBEA GeDEA-II GeDEA SPEA2 NSGA-II IBEA

(e) DTLZ5 boxplot (f) DTLZ6 boxplot

DTLZ7 test function

0.9
Hypervolume-Indicator

0.8

0.7

0.6

0.5

0.4

GeDEA-II GeDEA SPEA2 NSGA-II IBEA

(g) DTLZ7 boxplot

Figure 2.21: Box plots based on the Hypervolume metric. Each square contains six box plots
representing the distribution of Hypervolume values for the six algorithms. Results refer to
the DTLZ test suite.

on the preceding test functions. This is due to the fact that the variation range of
the decision variables was doubled when compared to the original one. Moreover,
the presence of the outlier, negligible from a statistical point of view, affects the
median value. As far as DTLZ1 test function is concerned, GeDEA-II is able to
reach the True Pareto Front unlike the competitors. Performance of GeDEA-II are
at the same level of those of the competitors on DTLZ2 test function, whereas it is
once again the only MOEA among those investigated able to reach the True Pareto
Front of the DTLZ3 test function. GeDEA-II shows similar behaviors to those of the
competitors on DTLZ4 and DTLZ4 test functions. Particularly remarkable are the
performance of our algorithm when the last two test functions of the DTLZ test
suite are considered. GeDEA-II is able to reach the True Pareto Fronts, whereas the
competitors remain trapped in the local Pareto Approximation Sets, as shown in
Figure2.19 and 2.20(a).
Finally, box plots prove, in general, that the performance of GeDEA-II are
superior to those of the competitors also as far as the repeatability of the results is
concerned.
52 2. GeDEAII: A powerful and robust MOEA

2.5 Conclusions
In this chapter the writer has presented GeDEA-II, an improved multi-objective
evolutionary algorithm that employs novel genetic operators compared to its prede-
cessor GeDEA as well as a new technique for adapt itself to different multiobjective
problems. Extensive numerical comparisons of GeDEA-II with GeDEA and with
NSGAII, SPEA-2 and IBEA, three state-of-the-art recently proposed algorithms,
have been carried out on various test problems. Moreover, optimization difficulties
have been enhanced further, in order to test the robustness of the codes. The key
results of the comparison are:

• GeDEA-II reaches the True Pareto fronts on all of the investigated test prob-
lems, while covering them in a satisfactory manner.

• GeDEA-II outperforms both its predecessor and the competitors on all the
test problems investigated.

• In extremely high dimensional spaces, GeDEA-II clearly shows unparalleled


performance over the other algorithms.

• Boxplots shows that the reproducibility of results of GeDEA-II is high-level,


when compared to that of the NSGAII, SPEA-2 and IBEA.

In addition to these characteristics, GeDEA-II performs these tasks with a reduced


number of objective functions evaluations, which is not negligible when considering
its application to real-world engineering problems. In chapter 7, two Test cases will
be considered, concerning mono and multi-objective problems where GeDEAII
will constitute the optimization engine. Results will demonstrate the performance
of GeDEAII.
Chapter 3

Parameterization techniques for


3D shape optimization in
aerodynamics

This chapter provides the basic classification of the most common shape defor-
mation techniques employed in the context of the aerodynamic shape optimization.
Finally, a brief introduction to Altair HyperMorph is presented, which is the mor-
phing tool chosen by AgustaWestland and the University of Padova to be used in
their optimization research program.

3.1 Shape parameterization techniques: An Overview


A survey of shape parameterization techniques is proposed in [78]. The most
used approches are as the following:

• Discrete Approach: it is based on using the coordinates of the boundary


points as design variables. This approach is easy to implement. However, it is
too much connected to the solution mesh and does not enable adaptive-mesh
flow evaluations. Due to the excessive refinement of the parametrization,
comparable to the resolution of the mesh, it is difficult to maintain a smooth
geometry of the optimized shape. For a model with a large number of grid
points, the number of design variables often becomes very large, which leads
to hight cost and a difficult optimization problem to solve. This technique al-
lows to easily parameterize complex 3D objects, but there are some problems
with definition of normals to a discrete surface. Hierarchical parameterization
using this approach might be done for example by agglomeration [60].

• Polynomial and Spline Approaches: the shape is described by a polynomial


curve in a very compact form with a small set of design variables . For
example, with a Bezier curve, the control points are used as design variables.
Despite recent progress, it is still difficult to parameterize and construct
complex, three-dimensional models based only on polynomial and spline

53
54 3. Parameterization techniques for 3D shape optimization in aerodynamics

representations. Complex shape requires a large number of control points.


The smoothness of the shape deformation is assured.

• CAD-based Approach: most solid modelling CAD systems use either a


boundary representation (B-Rep) or a constructive solid geometry method to
represent a physical, solid object. Based on a complete mathematical defini-
tion of a solid, it is possible to create a complete geometry. These systems
use Boolean operations such as intersection and union of simple features (e.g.
holes, slots, cuts, protrusion, fillets, chamfers, etc.). Existing Feature-based
solid modelling (FBSM) CAD tools are not capable of calculating sensitivity
derivatives analytically. After reconstructing the shape geometry, usually a
new mesh must be generated.

• Analytical Approach: the formulation is based on adding shape functions


(analytical functions) linearly to the baseline shape. All participating coef-
ficients are initially set to zero, so the first computation gives the baseline
geometry. The shape functions are smooth functions based on previous airfoil
design. This method results very good for wing parameterization, but not
simple to generalize for complex geometries.

• Free Form Deformation (FFD) Approach: FFD is one of the techniques


of deforming computer-generated objects, which comes from Computer
Graphics [80]. The FFD approach operates on the whole space regardless of
the representation of the deformed objects embedded in the space. Instead of
manipulating the object directly, FFD deforms a lattice that was built around
the object. The lattice is a space, in the shape of a cube or an arbitrary volume
[3], which wraps around the object. This lattice is basically a composite of
Bezier tensor patches in 3D called a Bezier volume, but it is also possible
to use B-spline or NURBS [55]. When we move the control points of the
lattice, the lattice is deformed. At the same time, the object inside the lattice
is deformed (see Figure 3.2).

3.2 The HyperMorph Parameterization Tool


For the purpose of this Thesis the geometry parameterization has been car-
ried out by means of the Altair HyperMorph morphing tool, implemented within
the Altair HyperMesh environment. This is a commercial software, distributed by
Altair Engineering, capable to assign parametric deformations to every kind of
finite elements model, allowing the designer to set up design of experiments and
optimization studies in a very simple way. HyperMorph provides a wide range
of warping and morphing techniques belonging to the first and the last groups
mentioned in section 3.1. See reference [3] to learn more about HyperMorph shape
deformation capabilities. In particular, as it will be see later on, the most effec-
tive methods for aerodynamic optimization purpose are the Domain-Handles and
the Morph-volume approaches, both belonging to the powerfull FFD technique
3.2 The HyperMorph Parameterization Tool 55
2.2 The HyperMorph Parameterization Tool

(a) Base model (b) Morphed model

Figure 2.1: Handles-Domain application.


Figure 3.1: Handles-Domain technique.

Regarding aerodynamic optimization matters, Handles-Domains approach is


the best way to generate a parametric model when small, local displacements are
required.
briefly discussed in the following sections. They will be both exploited during the
optimization works.
2.2.2 Morph-Volumes Approach
When the morph volumes approach is used, the model are enclosed in some
3.2.1 Domains-Handles Approachwhich
control volumes, the Morph-Volumes, are characterized by the presence
of Global Handles on their corners. When some of these handles are moved,
Whenthethe Domains-handles
nodes approach
and elements inside the is used,
corresponding the model
Morph-Volumes is divided into do-
are accordingly
mains where handles are used to control the relative domain shapes. When the
move in a really smooth manner.
The same prism, previously deformed with Domain-Handles (figure 2.1), is
handles now
are dealt
moved, the shape of the domains touching those handles change,
with Morph-Volumes (figure 2.2).
which in turn, changes the positions of the nodes inside those domains. During
the morphing process the mesh morphs in a logical way with nodes near the
moving handles moving more and nodes near the stationary handles moving less.
In the area between the handles, the mesh is stretched or compressed to match
the desired shape. An example of application of the Handles-Domains approach is
shown in Figure 3.1. In this example the prisms shown in figure are placed in a 3D
domain which is then (a)partitioned
Base model in six 3D domain, corresponding
(b) Morphed model to the six faces
of the prism, by the auto-partition algorithm [3].
Figure 2.2: Morph-Volumes This process generate the yellow
application.
handles on the corners of the prisms. Then some handles are manually placed in
In this example, two Morph-Volumes are created by the Pick on Screen tool
the middle section of the prism (Figure 3.1-a); the vertical displacement of these
[22]. The volumes can be visualized in figure 2.2 as the green regions around the
handles leads
prism. in
Thethe morphed
morphed model model shown
is obtained in Figure
by assign a vertical3.1-b.
displacement to the
As regards aerodynamic
central located, shape optimization, the Handles-Domains approach is
red, handles.
Because of the second order continuity in the deformed mesh achievable with
the best way to generate a parametric model when small, local displacements are
this morph-strategy, Morph-Volumes are really suited, for aerodynamic optimiza-
required,tion
and a precise
purpose, when control
global andofsmooth
the deformed region
shape changes has to be achieved.
are required.

3.2.2 Morph-Volume Approach


When the morph volumes approach is used, the model is enclosed 25 in some
control volumes, the Morph-Volumes, which are characterized by the presence of
Global Handles on their corners. When some of these handles are moved, the nodes
and elements inside the corresponding Morph-Volumes are accordingly moved in
a really smooth manner. The same prism, previously deformed with the Domain-
Handles (Figure 3.1), is now dealt with the Morph-Volumes (Figure 3.2). In this
example, two Morph-Volumes are created by means of the Pick on Screen tool [3].
The volumes can be visualized in Figure 3.2 as the green regions around the prism.
The morphed model is obtained by assigning a vertical displacement to the central
located, red, handles. Because of the second order continuity in the deformed
mesh achievable with this morph-strategy, Morph-Volumes are really suited, for
control volumes, the Morph-Volumes, which are characterized by the presence
of Global Handles on their corners. When some of these handles are moved,
the nodes and elements inside the corresponding Morph-Volumes are accordingly
move in a really smooth manner.
The same prism, previously deformed with Domain-Handles (figure 2.1), is
56 3. Parameterization techniques
now dealt with Morph-Volumes for 3D
(figure 2.2).shape optimization in aerodynamics

(a) Base model (b) Morphed model

Figure 2.2: Morph-Volumes application.


Figure 3.2: Morph-Volumes application technique.
In this example, two Morph-Volumes are created by the Pick on Screen tool
[22]. The volumes can be visualized in figure 2.2 as the green regions around the
prism. The morphed model is obtained by assign a vertical displacement to the
centraloptimization
aerodynamic located, red, handles.
purpose, when global and smooth shape changes are
Because of the second order continuity in the deformed mesh achievable with
required.this morph-strategy, Morph-Volumes are really suited, for aerodynamic optimiza-
tion purpose, when global and smooth shape changes are required.
3.2.3 Shapes as Design Variables
Into HyperMorph environment, shapes are collection of handles and/or nodes
perturbations. When a deformation on a given model is carried out, HyperMorph
25
stores the morph internally as a collection of perturbations. Saving a shape, ob-
tained by overlapping more simpler morphing operations, the handles and/or
nodes perturbations are stored in a new shape entity, which can be applied to
the undeformed model with any given scaling factor. Therefore, shapes created
in HyperMorph are suited to be use as parametric design variables in DOE and
optimization studies.
Chapter 4

Introduction to CFD

This chapter is intended as an introduction for Computational Fluid Dynamics


(CFD). Due to its introductory nature, only the basic principles of CFD are intro-
duced here. For a more detailed description, readers are referred to the precious
guide present in [94].

4.1 Introduction
Even though a univocal definition of CFD does not exist, here we can state
that CFD deals with the simulation of fluids engineering systems using modeling
(mathematical physical problem formulation) and numerical methods (discretiza-
tion methods, solvers, numerical parameters, and grid generations, etc.). CFD
provides a numerical approximation to the equations that govern fluid motion.
Application of the CFD to analyze a fluid problem requires the following steps.
First, the mathematical equations describing the fluid flow are written. These
are usually a set of partial differential equations.
These equations are then discretized to produce a numerical analogue of the
equations.
The domain is then divided into small grids or elements.
Finally, the initial conditions and the boundary conditions of the specific
problem are used to solve these equations. The solution method can be direct or
iterative. In addition, certain control parameters are used to control the convergence,
stability, and accuracy of the method.
All CFD codes contain three main elements:

• A pre-processor, which is used to input the problem geometry, generate the


grid, define the flow parameter and the boundary conditions to the code.

• A flow solver, which is used to solve the governing equations of the flow
subject to the conditions provided. There are four different methods used as
a flow solver:

* The finite difference method;


* The finite element method;

57
58 4. Introduction to CFD

* The finite volume;


* The spectral method.

• A post-processor, which is used to manage the data and show the results in
graphical and easy to read format.
In this chapter we are mainly concerned with the flow solver part of CFD, in
particular by paying attention on the system of equations to be solved. A brief
description of the RANS technique and of the finite-volume method will be finally
undertaken.

4.2 Governing equations


The Navier-Stokes equations are the basic governing equations for a viscous,
heat conducting fluid. It is a vector equation obtained by applying Newton’s
Law of Motion to a fluid element and is also called the momentum equation. It is
supplemented by the mass conservation equation, also called continuity equation
and the energy equation. Usually, the term Navier-Stokes equations is used to refer
to all of these equations. In addition to these equations, further equation can be
solved, like, for example, the equations of state for real or ideal gas, cavitation
models, magnetohydrodynamics models, and so on.
However, in this chapter, only the Navier-Stokes equations will be considered.

4.2.1 The Continuity Equation


The equations governing the fluid motion are the three fundamental principles
of mass, momentum, and energy conservation. Let start with the first principle,
which states that
...the total amount of mass inside any region can only change by the amount that passes
in or out of the region through the boundary.
One considers the infinitesimal control volume depicted in Figure 4.1.
The law of conservation of mass states that
dm
dt
= ∑ ṁ − ∑ ṁ (4.1)
in out

In the three-dimensional space, the Eq. 4.1 assumes the following form:
∂ρ∆x∆y∆z
= (ρ u)∆y∆z + (ρ v)∆x∆z + (ρ w)∆x∆y
∂t
∂(ρu)
 
− (ρ u) + ∆x ∆y∆z
∂x
(4.2)
∂(ρv)
 
− (ρ v) + ∆y ∆x∆z
∂y
∂(ρw)
 
− (ρ w) + ∆z ∆x∆y
∂z
4.2 Governing equations 59

Figure 4.1: An infinitesimal fluid control volume - The Continuity equation.

which, after some rearrangements is given the following form:


∂ρ ∂(ρu) ∂(ρv) ∂(ρw)
+ + + = 0. (4.3)
∂t ∂x ∂y ∂z
When the flow is at steady-state, ρ does not change with respect to time. The
continuity equation is reduced to:
∂(ρu) ∂(ρv) ∂(ρw)
+ + = 0. (4.4)
∂x ∂y ∂z
When the flow is incompressible, ρ is constant and does not change with respect
to space. The continuity equation is reduced to:
∂u ∂v ∂w
+ + = 0. (4.5)
∂x ∂y ∂z

By using the substantive derivative8 concept, the mass continuity equation assumes
the general form:

+ ρ(∇ · v) = 0. (4.6)
Dt
8 The substantial derivative is a derivative taken along a path moving with velocity v, and is often
used in fluid mechanics and classical mechanics. It describes the time rate of change of some quantity
(such as heat or momentum) by following it, while moving with a space- and time-dependent velocity
field. For example, in fluid dynamics, take the case that the velocity field under consideration is the
flow velocity itself, and the quantity of interest is the temperature of the fluid. Then the material
derivative describes the temperature evolution of a certain fluid parcel in time, as it is being moved
along its pathline (trajectory) while following the fluid flow. The total or substantial derivative assumes
the following mathematical expression, for a scalar, ϕ, on the left, or a vector, u, on the right:
Dϕ ∂ϕ Du ∂u
= + v · ∇ ϕ, = + v · ∇u
Dt ∂t Dt ∂t
60 4. Introduction to CFD

4.2.2 The Momentum Equation


The principle of momentum conservation comes directly from the Newton’s
second law, which states that

...the net force on a particle is equal to the time rate of change of its linear momentum p
in an inertial reference frame.
Introduzione alla CFD 1.4 Equazioni che governano il
is given the following mathematical expression:
IL METODO DEI VOLUMI FINITI 2a legge
~F = m ~a (4.7)
r r
Seconda legge di Newton: F = ma
Let one apply the principle stated in Eq. 4.7 to the infinitesimal, space fixed control
volume depicted La in Figureora
si applichi 4.2.
al The forces
volume to be (infinitesimo
di controllo considerede fisso
are the superficial forces risulta
in dire
nello spazio) di seguito considerato.

Sommando le precedenti e dividendo per i

risultante sfor
agenti lungo x
volume)

La seconda legge di Newton lungo x diven

Analogamente, nelle altre due direzioni:

Le forze in gioco sono forze di superficie (normali e tangenziali)


Figure 4.2: An infinitesimal fluid control volume - The Momentum Equation.
e forze di volume (come la forza di gravità). Lungo la direzione
x si ha:
(normal and tangential to the body) and the volume forces, such as the gravity
forces. The Eq. 4.8 and 4.9 represent the overall normal
risultante and tangential stresses,
sforzi normali
in direzione x
respectively, acting on the control volume depicted Figure 4.2, along x-direction.
 
∂σxx
σxx + ∆x ∆y∆z − σxx ∆y∆z overall normal stresses acting on
∂x (4.8)
the control volume along x-direction
 
∂τyx
τyx + ∆y ∆x∆z − τyx ∆x∆z overall tangential stresses acting on
∂y
  (4.9)
∂τzx
τzx + ∆z ∆x∆y − τzx ∆x∆y the control volume along x-direction
∂z
By adding the terms in Eq. 4.8 and 4.9 and dividing by the volume, one can obtain

∂σxx ∂τyx ∂τzx


+ + results of the superficial stress
∂x ∂y ∂z (4.10)
along x-direction (per volume unit)
4.2 Governing equations 61

The Newton’s second law, evaluated along x-direction, becomes

Du ∂σxx ∂τyx ∂τzx


+ ∑ Fx
body f orces
ρ = + + (4.11)
Dt ∂x ∂y ∂z

Similarly, along y-direction and y-direction the conservation of momentum gives,


respectively,

Dv ∂τxy ∂σyy ∂τzy


+ ∑ Fy
body f orces
ρ = + +
Dt ∂x ∂y ∂z
(4.12)
Dw ∂τxz ∂τyz ∂σzz
+ ∑ Fz
body f orces
ρ = + +
Dt ∂x ∂y ∂z

Consider now a Newtonian9 and isotropic10 fluid. Thanks to this assumption,


it can be stated that

σxx = − p + τxx
σyy = − p + τyy
σzz = − p + τzz
∂u ∂u ∂v ∂w
τxx = 2µ +λ + +
∂x ∂x ∂y ∂z
∂v ∂u ∂v ∂w
τyy = 2µ + λ + +
∂y ∂x ∂y ∂z
∂z ∂u ∂v ∂w (4.13)
τzz = 2µ + λ + +
∂z ∂x ∂y ∂z
 
∂v ∂u
τxy = τyx = µ +
∂x ∂y
 
∂w ∂u
τxz = τzx = µ +
∂x ∂z
 
∂w ∂v
τyz = τzy = µ +
∂y ∂z

λ is the second coefficient of viscosity (related to bulk viscosity). The value of λ,


which produces a viscous effect associated with volume change, is very difficult to
determine, not even its sign is known with absolute certainty. Even in compressible
flows, the term involving λ is often negligible; however it can occasionally be
important even in nearly incompressible flows and is a matter of controversy.
When taken nonzero, the most common approximation is λ ≈ −2/3µ [12].

9ANewtonian fluid (named after Isaac Newton) is a fluid whose shear stress τ versus strain rate
∂u
∂y curve is linear and passes through the origin. The constant of proportionality is known as the
dynamic viscosity µ.
10 A fluid whose properties are not dependent on the direction along which they are measured.
62 4. Introduction to CFD

By substituting relations given in 4.13, in Eq. 4.11 and 4.12, one can obtain the
Navier-Stokes equations.
  
Du ∂p ∂ ∂u ∂u ∂v ∂w
ρ =− + 2µ +λ + +
Dt ∂x ∂x ∂x ∂x ∂y ∂z
     
∂ ∂u ∂v ∂ ∂u ∂w
+ ∑ Fx
body f orces
µ + + µ +
∂y ∂y ∂x ∂z ∂z ∂x
  
Dv ∂p ∂ ∂v ∂u ∂v ∂w
ρ =− + 2µ + λ + +
Dt ∂y ∂y ∂y ∂x ∂y ∂z
      (4.14)
∂ ∂u ∂v ∂ ∂v ∂w
+ ∑ Fy
body f orces
µ + + µ +
∂x ∂y ∂x ∂z ∂z ∂y
  
Dw ∂p ∂ ∂w ∂u ∂v ∂w
ρ =− + 2µ +λ + +
Dt ∂z ∂z ∂z ∂x ∂y ∂z
     
∂ ∂u ∂w ∂ ∂v ∂w
+ ∑ Fz
body f orces
µ + + µ +
∂x ∂z ∂x ∂y ∂z ∂y
or, after accounting for the gravity as a body force, and developing the substan-
tial derivatives,

 
∂u ∂u ∂u ∂u ∂p
ρ +u +v +w =−
∂t ∂x ∂y ∂z ∂x
 2
∂ u ∂2 u ∂2 u

+µ + 2 + 2 + ρgx
∂x2 ∂y ∂z
 
∂v ∂v ∂v ∂v ∂p
ρ +u +v +w =−
∂t ∂x ∂y ∂z ∂y
 2 (4.15)
∂ v ∂2 v ∂2 v

+µ + 2 + 2 + ρgy
∂x2 ∂y ∂z
 
∂w ∂w ∂w ∂w ∂p
ρ +u +v +w =−
∂t ∂x ∂y ∂z ∂z
 2
∂ w ∂2 w ∂2 w

+µ + 2 + 2 + ρgz .
∂x2 ∂y ∂z
Finally, the vector form of the Momentum equation is presented in Eq.4.16, in
order to draw a parallel with the Eq. 4.6
 
∂v
ρ + v · ∇v = −∇ p + µ∇2 v + (λ + µ)∇v + f. (4.16)
∂t

4.2.3 The Energy Equation


The principle of Energy conservation comes directly from the First Thermody-
namic law, which states that
There is a state function E, called “energy”, whose differential equals the work exchanged
with the surroundings during an adiabatic process.11
11 This definition was given by Rudolf Clausius in 1850.
4.2 Governing equations 63

Mathematically, one writes


∆E = W + Q (4.17)
where E is the total energy of the system, whereas Q and W are the amounts
of heat supplied to the system and work done by the system, respectively.
Let one consider the contribution given by work, W, in Eq. 4.17, applied to the
infinitesimal, space fixed control volume depicted in Figure 4.3.

Figure 4.3: An infinitesimal fluid control volume. The Work contribution to the Energy
equation.

Let one consider both work and energy terms, separately. The forces considered
are the superficial forces (normal and tangential to the body). Work done (per time
unit) by the normal and tangential stresses acting on the control volume depicted
Figure 4.3, along x-direction, is given in Eq. 4.18 and 4.19, respectively.

 
∂uσxx
uσxx + ∆x ∆y∆z − uσxx ∆y∆z Work done by the normal stresses acting on
∂x
the control volume along x-direction
(4.18)
 
∂uτyx
uτyx + ∆y ∆x∆z − uτyx ∆x∆z Work done by the tangential stresses acting on
∂y
 
∂uτzx
uτzx + ∆z ∆x∆y − uτzx ∆x∆y the control volume along x direction
∂z
(4.19)
By adding the terms in Eq. 4.18 and 4.19 and dividing by the volume, one can
obtain
∂uσxx ∂uτyx ∂uτzx
+ + Resulting Work done by the superficial stress
∂x ∂y ∂z (4.20)
along x-direction (per volume and time unit)

Similarly, along y-direction and y-direction, the Work done by superficial forces
is, respectively,
64 4. Introduction to CFD

∂vτxy ∂vσyy ∂vτzy


+ + Resulting Work done by the the superficial stress
∂x ∂y ∂z (4.21)
along y-direction (per volume and time unit)

∂wσxz ∂wτyz ∂wσzz


+ + Resulting Work done by the the superficial stress
∂x ∂y ∂z (4.22)
along z-direction (per volume and time unit)

Now, let us consider the contribution given by Heat, Q, by referring to the


infinitesimal control volume depicted in Figure 4.4.

Figure 4.4: An infinitesimal fluid control volume. The Heat contribution to the Energy
equation.

The net Heat flux, along x-direction, is

 
∂q x
qx + ∆x ∆y∆z − q x ∆y∆z Heat flux along x-direction
∂x (4.23)

Similarly, along y-direction and z-direction, the Heat flux is, respectively,

 
∂qy
qy + ∆y ∆x∆z − qy ∆x∆z Heat flux along y-direction
∂y (4.24)

 
∂qz
qz + ∆z ∆x∆y − qz ∆x∆y Heat flux along z-direction
∂z (4.25)

By adding the terms in Eq. 4.23, 4.24 and 4.25, and dividing by the volume, one
can obtain the overall heat flux crossing control volume in Figure 4.4:

∂q x ∂qy ∂qz
+ + (4.26)
∂x ∂y ∂z
4.2 Governing equations 65

By gathering all the terms in Eqs. 4.20, 4.21, 4.22 and 4.26, the First Thermody-
namic law, can be stated as:

DE ∂uσxx ∂vσyy ∂wσzz ∂uτyx ∂uτzx ∂vτxy


ρ = + + + + +
Dt ∂x ∂y ∂z ∂y ∂z ∂x
(4.27)
∂vτzy ∂wτxz ∂wτyz ∂q x ∂qy ∂qz
+ + + − − −
∂z ∂x ∂y ∂x ∂y ∂z
By considering the Fourier’s Law of Conduction for the thermal fluxes:
∂T ∂T ∂T
q x = −k qy = −k qz = −k , (4.28)
x y z
where k is the referred to as thermal conductivity, and by resolving normal stresses
as indicated in Eq. 4.13, one can finally obtain:
     
DE ∂ ∂T ∂ ∂T ∂ ∂T
ρ = k + k + k
Dt ∂x x ∂y y ∂z z
(4.29)
∂up ∂vp ∂wp
− − − +Φ
∂x ∂y ∂z
Φ collects the following terms and takes into account the energy dissipation
due to the fluid viscosity (that is, the heat created because of friction, at the expense
of mechanical energy).

∂uτxx ∂uτyx ∂uτzx


Φ= + +
∂x ∂y ∂z
∂vτxy ∂vτyy ∂vτzy
+ + + (4.30)
∂x ∂y ∂z
∂wτxz ∂wτyz ∂wτzz
+ + +
∂x ∂y ∂z
Now, it is worth noting that for a compressible flow, it can be stated that

p 1 2 p
u + v2 + w2 = E + (4.31)

h = e+ +
ρ 2 ρ
where h is the fluid total enthalpy, whereas e is the internal energy.
By neglecting kinetic contribution,

h = hst = c p T (4.32)

where c p is the specific heat capacity at constant pressure, not temperature


dependent.
Finally, let us consider the special case, where the fluid is incompressible and
the continuity equation applies. As we said before, by neglecting the kinetic energy
the enthalpy h can be reduced to CpT, where Cp is the specific heat and is assumed
to be constant. Equation 4.29 can be expressed as
66 4. Introduction to CFD

     
DT ∂ ∂T ∂ ∂T ∂ ∂T ∂p
ρCp = k + k + k + +Φ (4.33)
Dt ∂x x ∂y y ∂z z ∂t

4.3 Turbulence Modeling


There are two radically different states of flows that are easily identified and
distinguished: laminar flow and turbulent flow.
Laminar flows are characterized by smoothly varying velocity fields in space
and time in which individual “laminae” (sheets) move past one another without
generating cross currents. These flows arise when the fluid viscosity is sufficiently
large to damp out any perturbations to the flow that may occur due to boundary
imperfections or other irregularities. These flows occur at low-to-moderate values
of the Reynolds number.
In contrast, turbulent flows are characterized by large, nearly random fluctua-
tions in velocity and pressure in both space and time. These fluctuations arise from
instabilities that grow until nonlinear interactions cause them to break down into
finer and finer whirls that eventually are dissipated (into heat) by the action of
viscosity. Turbulent flows occur in the opposite limit of high Reynolds numbers. In
Figure 4.5, the laminar-to-turbulent transition over a submarine hull is depicted12 .
In the upstream region the flow is laminar, but soon after the transition to tur-
bulent flow occurs. One of the most important
20091226210527!USS_Los_Angeles;0868802.jpg (...
characteristics of turbulent flow
http://upload.wikimedia.org/wikipedia/commons/ar...

Figure 4.5: Laminar-to-turbulent transition over a submarine hull (image taken from the
NavSource Naval History website).

concerns its dissipative behaviour, when compared to that characterizing laminar


flow. In Figure 4.6, the skin friction coefficient of a NACA 23012 airfoil is plotted.
Results are obtained with XFOIL CFD code13 . The sudden increase of the friction
12 Please visit the website http://www.navsource.org/archives/08/08688.htm
13 XFOIL is an interactive program for the design and analysis of subsonic isolated airfoils. Please
visit website http://web.mit.edu/drela/Public/web/xfoil/
1 di 1 29/12/2010 11:51
4.3 Turbulence Modeling
Introduzione alla CFD 67
IL METODO DEI VOLUMI FINITI Outline

coefficient Flusso
onceturbolento:
the transition
più dissipativo is occurred
del flusso laminare is clearly visible.

Upper Surface

Laminar to
turbulent
transition Lower Surface

XFOIL Output

Figure 4.6: Skin friction coefficient of NACA 23012 airfoil. XFOIL output

Another distinctive feature of turbulent flows is that turbulent boundary layer


is more “energetic” than the laminar one. Thanks to this characteristics, turbulent
flows can withstand higher adverse pressure gradients than those sustainable
by laminar flows. This aspect of turbulent flows is exploited in devices such as
vortex generator or synthetic jets, which will be described in Sections 5.5.2 and
5.5.1, respectively. Possible applications of such devices regard turbomachinery
aerodynamics (to postpone or delay flow separation means increased performances)
and external aerodynamics (to postpone or delay flow separation means reduced
drag and, in general, better flight performances).
Let one reconsider now the laminar-to-turbulent transition phenomenon intro-
duced above and look at Figure 4.7, which shows an induced transition experiment.
A typical time history of the flow variable u at a fixed point in space is shown
in the lower side of Figure 4.7. The dashed line through the curve indicates the
“average” velocity. Instantaneous velocity can be defined as:

u(t) = ū + u0 (t) (4.34)

where u0 (t) represents the fluctuating part, whereas ū is the average part.
Two questions now arise:

1. Ho can we numerically compute any fluid-dynamics turbulent flow field?

2. In the case that turbulence cannot be directly computed due to computational limi-
tations, how can we take into account its effects?

In order to answer to these fundamental questions, section 4.3.1 will be devoted


to a widely used technique, that is the RANS technique, able to take into account
the effects of the fluctuating part in Eq. 4.34, without directly solve it.
Introduzione alla CFD
68 IL METODO DEI VOLUMI FINITI 4. Introduction
2.1 Cos’è to CFD (4/4)
la turbolenza?

Laminar to turbulent transition

Induced
instability

u (t ) = u + u ᄁ
(t )

Figure 4.7: Laminar to turbulent induced transition.

4.3.1 The RANS Technique


In studying turbulent flows, the objective is to obtain a theory or a model that
can yield quantities of interest, such as velocities. For turbulent flow, the range of
length scales and complexity of phenomena make most approaches impossible.
The primary approach in this case is to create numerical models to calculate the
properties of interest. A selection of some commonly-used computational models
for turbulent flows are presented in this section.
The chief difficulty in modelling turbulent flows comes from the wide range
of length and time scales associated with turbulent flow. As a result, turbulence
models can be classified based on the range of these length and time scales that
are modeled and the range of length and time scales that are resolved. The more
turbulent scales that are resolved, the finer the resolution of the simulation, and
therefore the higher the computational cost. If a majority or all of the turbulent
scales are modeled, the computational cost is very low, but the tradeoff comes in
the form of decreased accuracy.
In addition to the wide range of length and time scales and the associated
computational cost, the governing equations of fluid dynamics contain a non-linear
convection term and a non-linear and non-local pressure gradient term. These
nonlinear equations must be solved numerically with the appropriate boundary
and initial conditions. The three main approaches commonly used to manage
turbulence in CFD calculations are:
• Direct Numerical Simulation (DNS): Direct numerical simulation resolves
the entire range of turbulent length scales. This means that the whole range
of spatial and temporal scales of the turbulence must be resolved. All the
spatial scales of the turbulence must be resolved in the computational mesh,
Direct Numerical Simulation http://www.ecs.umass.edu/mie/faculty/debk/Resear...

4.3 Turbulence Modeling 69


Direct Numerical Simulation of
Turbulent
from Flows scales (Kolmogorov scales), up to the integral
the smallest dissipative
scale L, associated
To study with
the details of the motions
a turbulent flow, it iscontaining
sometimes more most of the tokinetic energy.
informative
This marginalizes
accurately theflow
simulate the effect
with aof models,
computer thanbut istoextremely
to try observe it in expensive.
the The
laboratory. Direct numerical simulation involves3the numerical solution of the
computational cost isfluid
equations that govern proportional to Re [68].
flows. It is a research Hence,
tool that only
provides lowan− Reynolds
us with
flows can be studied with this technique, that are not so common
extremely detailed description of the flow field. Expand the images below to in practical
see the fine structure that occurs when the red and blue fluids mix and react
engineering problems.
in a turbulent DNSonisthe
flow. The image intractable for flows
left shows where with complex
the reaction is geometries
or flow configurations. Figure 4.8 shows two fluids reacting in a turbulent
occurring, while the one on the right shows how the two fluids have mixed.
They are taken from a DNS with 512x512x1024 grid points run at the Arctic
flow. TheSupercomputing
Region image on theCenter.left, shows where the reaction is occurring, while the
one on the right shows how the two fluids have mixed.

Figure 4.8:
TheDirect
images Numerical Simulation
below are slices from 256 of3 DNSs
two reacting
in which flows
a blob (image available
of fuel reacts with at website
http://www.ecs.umass.edu/mie/faculty/debk/Research/dns.html
oxidant in a one-step, irreversible reaction. The flow is incompressible,). and the
fluid properties are constant. The plots show that the reaction zones are thin,
even when the reaction is isothermal, and become thinner when the reaction
rate depends on temperature. Because the reaction zones are so thin,
• Large eddytechiques
modeling simulation (LES):
involving Largewhich
averaging, eddyare
simulation
common foris non-reacting
a technique in which
theturbulence,
smallest cannot
scalesbeofused.
the flow are removed through a filtering operation,
and their effect modelled using subgrid scale models. This allows the largest
Mixture Fraction Isothermal Reaction Rate
and most important scales of the turbulence to be resolved, while greatly
reducing the computational cost incurred by the smallest scales. This method
requires greater computational resources than RANS methods,
1 di 3
but is far
29/12/2010 15:09

cheaper than DNS.


LESPremixedFlame.jpg Figure
(Immagine 4.9 shows
JPEG, 776x275 ... anhttp://upload.wikimedia.org/wikipedia/commons/6/...
example of LES application.

Figure 4.9: Volume rendered image of a Large Eddy Simulation of a non-premixed swirl
flame (taken from [86]).

• Reynolds-averaged Navier-Stokes (RANS): Reynolds-averaged Navier-Stokes


equations are the oldest approach to turbulence modeling. An ensemble
version of the governing equations is solved, which introduces new apparent
70 4. Introduction to CFD

stresses known as Reynolds stresses. This adds a second order tensor of


unknowns for which various models can provide different levels of closure. To
put it simply, neither the smallest scale nor the largest ones of the turbulence
6. Results and discussion
are resolved, but its macroscopic effects are taken into account anyhow. It is a
common misconception that the RANS equations do not apply to flows with a
time-varying mean flow because these equations are “time-averaged”. In fact,
statistically unsteady (or non-stationary) flows can equally be treated. This
is sometimes referred to as URANS. There is nothing inherent in Reynolds
averaging to preclude this, but the turbulence models used to close the
equations are valid only as long as the time over which these changes in
the mean occur is large compared to the time scales of the turbulent motion
containing most of the energy. In Figure 4.10 an application of the unsteady-
RANS technique is presented, regarding turbomachinery flow.

Figure 4.10: Viscous wake impinging a stator vane row of an aero-engine LP turbine. (Axial
velocity perturbation magnitude coloured)(taken from [20])

Since the RANS approach implemented in the CFD codes Fluent will be exploited
in the following introduced optimization loop, it is now deeply discussed.
RANS models can be divided into two broad approaches:
* Boussinesq hypothesis14 : This method involves using an algebraic equation
for the Reynolds stresses which include determining the turbulent viscosity,
and depending on the level of sophistication of the model, solving transport
equations for determining the turbulent kinetic energy κ and dissipation (e,
14 Joseph Valentin Boussinesq (13 March 1842 - 19 February 1929) was a French mathematician and

physicist who made significant contributions to the theory of hydrodynamics, vibration, light, and
heat.
4.3 Turbulence Modeling 71

or turbulent dissipation rate in the κ − e model [56], ω, or specific dissipation


rate in the κ − ω model [61]). The models available in this approach are often
referred to by the number of transport equations associated with the method.
For example, the Mixing Length model is a “Zero Equation” model because
no transport equations are solved; the κ − e is a “Two Equation” model
because two transport equations (one for κ and one for e) are solved.

* Reynolds stress model (RSM): This approach attempts to actually solve


transport equations for the Reynolds stresses. This means introduction of
several transport equations for all the Reynolds stresses and hence this
approach is much more costly in CPU effort.

Let consider now the 2-D version of continuity equation expressed in Eq. 4.5
for incompressible flows. It assumes the following expression:

∂u ∂v
+ = 0. (4.35)
∂x ∂y

At this point, let substitute the expression of u given in Eq. 4.34 in each velocity
term of the previous equation, that is, let substitute each instantaneous quantity
into its averaged and fluctuating components fields.
One obtains
∂ (ū + u0 (t)) ∂ (v̄ + v0 (t))
+ = 0. (4.36)
∂x ∂y
Time-averaging this equation yields,

∂ (ū + u0 (t)) ∂ (v̄ + v0 (t)) ∂ (ū + u0 (t)) ∂ (v̄ + v0 (t))


+ = + =0 (4.37)
∂x ∂y ∂x ∂y

It is worth noting that the temporal mean of the velocity fluctuating part is
Z t+ T h i
u0 (t) = lim u(t) − u dt = 0 (4.38)
T →∞ t

Moreover, the following properties inherent to time-averaging are to be taken


into account:

∂u ∂u ∂u0 (t) ∂u
= + =
∂x ∂x ∂x ∂x (4.39)
uv = (ū + u0 (t)) + (v̄ + v0 (t)) = ūv̄ + u0 (t)v0 (t)

Eq. 4.37 becomes finally,

∂u
∂x + ∂v
∂y =0 (4.40)

Similarly, momentum equation and energy equation become respectively, after


time-averaging
72 4. Introduction to CFD

   
1 ∂ p̄ ∂ū ∂v̄
∂u
∂t+ ∂uu + ∂uv
= − + ∂
ν + ∂
ν
 ∂x ∂y
ρ ∂x h ∂x ∂x ∂y ∂y
∂[u0 (t)u0 (t)] ∂[u0 (t)v0 (t)]
 i
∂ū ∂ū
+ ∂x ν ∂x + ∂y ν ∂y −
∂ ∂
∂x + ∂y

    (4.41)
∂v
∂t + ∂uv + ∂vv
= − 1ρ ∂∂yp̄ + ∂x
∂ ∂v̄
ν ∂x + ∂y∂ ∂v̄
ν ∂y
 ∂x ∂y
 h 0 0
∂[v0 (t)v0 (t)]
 i
∂[u (t)v (t)]
+ ∂x

ν ∂∂xū + ∂
∂y ν ∂ū
∂y − ∂x + ∂y

and

1 ∂ p̄
     
∂v ∂uv ∂vv ∂ ∂v̄ ∂ ∂v̄ ∂ ∂ū
+ + =− + ν + ν + ν
∂t ∂x ∂y ρ ∂y ∂x ∂x ∂y ∂y ∂x ∂x
 " # (4.42)
∂[u0 (t)v0 (t)] ∂[v0 (t)v0 (t)]

∂ ∂ū
+ ν − +
∂y ∂y ∂x ∂y

At this point, closing the previous RANS equations requires modeling the
Reynold’s stress Rij 15
Joseph Boussinesq was the first practitioner of this, introducing the concept of
eddy viscosity. The Boussinesq assumption states that the Reynolds stress tensor,
Rij , is proportional to the mean strain rate tensor, Sij∗ , and can be written in the
following way:
2
Rij = 2 µt Sij∗ − ρkδij (4.43)
3
where µt is a scalar property called the eddy viscosity. The same equation can
be written more explicitly as:

∂u 2 ∂ρ
−ρu0 u0 = 2µt −
∂x 3 k
∂v 2 ∂ρ
−ρv0 v0 = 2µt − (4.44)
∂y 3 k
 
0 0
∂v ∂v
−ρu v = µt +
∂x ∂y

The right-hand side is analogous to Newton’s law of viscosity, except for the appear-
ance of the turbulent or eddy viscosity µt and turbulent kinetic energy κ.
In Eq. 4.44, the turbulent momentum transport is assumed to be proportional
to the mean gradients of velocity. Similarly, the turbulent transport of temperature
is taken to be proportional to the gradient of the mean value of the transported
quantity. In other words,
15 The non-linear term υi0 υ0j from the convective acceleration is known as the Reynolds stress,

Rij = υi0 υ0j


4.3 Turbulence Modeling 73

∂T
−ρu0 T 0 = Γt
∂x
(4.45)
∂T
−ρv0 T 0 = Γt
∂y

where κ = 21 u0 u0 is the turbulent kinetic energy whereas Γt = PrTT is referred to


µ

as the turbulent thermal diffusivity, Pr T being the turbulent Prandtl number.


By substituting the Reynolds stress expressions in Eq. 4.44 and the extra
temperature transport terms in Eq. 4.45 into the governing Eqs. 4.40, 4.41, and 4.42,
and removing the overbar, that is by default indicating the average quantities, we
obtain

∂u
∂x + ∂v
∂y =0 (4.46)

h i h i
∂u
∂t + ∂uu
∂x + ∂uv
∂y = − 1ρ ∂p
∂x +

∂x (ν + νT ) ∂u
∂x i + ∂
∂y h ( ν + ν ) ∂v
T ∂y
h i
+ ∂x

(ν + νT ) ∂x + ∂y (ν + νT ) ∂u
∂u ∂
∂y

h i h i (4.47)
∂v
∂t + ∂uv
∂x + ∂vv
∂y = − 1ρ ∂p
∂y +

∂x (ν + νT ) ∂x
∂v
+ ∂y

(ν + νT ) ∂v
∂y
h i h i
+ ∂x

(ν + νT ) ∂u
∂x + ∂
∂y ( ν + ν ) ∂u
T ∂y

1 ∂p
     
∂v ∂uv ∂vv ∂ ν νT ∂v ∂ ν νT ∂v
+ + =− + + + +
∂t ∂x ∂y ρ ∂y ∂x Pr Pr T ∂x ∂y Pr Pr T ∂y
     
∂ ν νT ∂u ∂ ν νT ∂u
+ + + +
∂x Pr Pr T ∂x ∂y Pr Pr T ∂y
(4.48)

Eq. 4.46, 4.47 and 4.48 constitute respectively the Time averaged Continuity equa-
tion, the Time averaged Momentum equation and the Time averaged Energy equation.

4.3.2 The κ − ω turbulence model


In the previous time-averaged equations a new variable appears, that is the
kinematic eddy viscosity νT . In order to make it possible to solve this system of
equations, one or more additional equation are to be written, which constitute the
“Turbulence models”.
In this section, the κ − ω model will be addressed, since it is one of the most
common turbulence models and it will be exploited during the CFD simulations
reported in the chapter 7. It is a two equation model, that means, it includes two
extra transport equations to represent the turbulent properties of the flow. This
74 4. Introduction to CFD

allows a two equation model to account for history effects like convection and
diffusion of turbulent energy.
The first transported variable is turbulent kinetic energy, κ. The second trans-
ported variable in this case is the specific dissipation, ω. It is the variable that
determines the scale of the turbulence, whereas the first variable, κ, determines
the energy in the turbulence.
In particular, the SST (acronym for Shear Stress Transport) κ − ω turbulence
model [61] is a two-equation eddy-viscosity model which has become very popular.
The shear stress transport (SST) formulation combines the best of two worlds. The
use of a κ − ω formulation in the inner parts of the boundary layer makes the
model directly usable all the way down to the wall through the viscous sub-layer,
hence the SST κ − ω model can be used as a Low-Re turbulence model without any
extra damping functions. The SST formulation also switches to a κ − e behaviour
in the free-stream and thereby avoids the common κ − ω problem that the model
is too sensitive to the inlet free-stream turbulence properties. Authors who use the
SST κ − ω model often merit it for its good behavior in adverse pressure gradients
and separating flow. The SST κ − ω model does produce a bit too large turbulence
levels in regions with large normal strain, like stagnation regions and regions with
strong acceleration. This tendency is much less pronounced than with a normal
κ − e model though.
The kinematic eddy viscosity is calculated as

a1 k
νT = (4.49)
max( a1 ω, SF2 )

Turbulence Kinetic Energy is modeled via the following equation:

 
∂k ∂k ∗ ∂ ∂k
+ Uj = Pk − β kω + (ν + σk νT ) (4.50)
∂t ∂x j ∂x j ∂x j

Specific Dissipation Rate transport equation is

 
∂ω ∂ω 2 2 ∂ ∂ω
+ Uj = αS − βω + (ν + σω νT )
∂t ∂x j ∂x j ∂x j
(4.51)
1 ∂k ∂ω
+ 2(1 − F1 )σω2
ω ∂xi ∂xi

Closure Coefficients and Auxiliary Relations are


4.4 Generic form of the Governing Equations for CFD 75

" √ !#2 
2 k 500ν
F2 = tanh  max ,
β∗ ωy y2 ω

 
∂Ui
Pk = min τij , 10β∗ kω
∂x j
( " √ ! #)4 
 k 500ν 4σω2 k 
F1 = tanh min max , ,
 β∗ ωy y2 ω CDkω y2 

1 ∂k ∂ω
 
CDkω = max 2ρσω2 , 10−10
ω ∂xi ∂xi (4.52)
φ = φ1 F1 + φ2 (1 − F1 )
5
α1 = , α2 = 0.44
9
3
β1 = , β 2 = 0.0828
40
9
β∗ =
100
σk1 = 0.85, σk2 = 1
σω1 = 0.5, σω2 = 0.856

4.4 Generic form of the Governing Equations for CFD


From the governing equations derived above, there are significant common-
alities between these various equations. If we introduce a general variable Φ
expressing all the fluid flow equations, including equations of temperature and
turbulent quantities, in the conservative incompressible form, the equation can
usually be written as

∂ρΦ
+ ∇ · (ρ~uΦ) = ∇ · (Γ∇Φ) + SΦ (4.53)
∂t } | {z } | {z } |{z}
Source term
| {z
Convection term Di f f usion term
Transient term
where Γ is the diffusion coefficient or diffusivity.

• The transient term, ∂t , accounts for the accumulation of Φ in the concerned


∂ρΦ

control volume;

• The convection term, ∇ · (ρ~uΦ), accounts for the transport of Φ due to the
existence of the velocity field (note the velocity ~u multiplying Φ);

• The diffusion term, ∇ · (Γ∇Φ), accounts for the transport of Φ due to its
gradients;

• The source term, SΦ , accounts for any sources or sinks that either create
or destroy Φ. Any extra terms that cannot be cast into the convection or
diffusion terms are considered as source terms.
76 4. Introduction to CFD

After developing all the terms in eq. 4.53, we can obtain

 
∂Φ ∂uΦ ∂vΦ ∂wΦ ∂ ∂Φ
+ + + = Γ
∂t ∂x ∂y ∂z ∂x ∂x
    (4.54)
∂ ∂Φ ∂ ∂φ
+ Γ + Γ + SΦ
∂y ∂y ∂z ∂z

Equation 4.54 is the so-called Transport equation for the property Φ.


The goal of all discretization techniques (Finite Difference, Finite Element,
Finite Volume, Boundary Element...) is to derive a mathematical formulation to
transform each of these terms into an algebraic equation. Once applied to all
control volumes in a given mesh, and after setting the transport property φ equal
to ρ u, v, w, T, k, E, and selecting appropriate values for the diffusion coefficient Γ
and source terms Sφ we obtain a full linear system of equations that needs to be
solved.

4.5 The Finite Volume method: an overview


In order to discretize in the physical spaces the integral form of the conservation
equations, whose general form is expressed in Eq. 4.54, the Finite-Volume method
method is required.
The Finite Volume Method (FVM) is one of the most versatile discretization
techniques used in CFD.
The computational domain is first subdivided into a finite number of contigu-
ous control volumes, where the resulting statements express the exact conservation
of relevant properties for each of the control volumes. At the centroid of each of
the control volumes, the variable values are then calculated. To do that, integration
of the differential form of the governing equations over each control volume is
required. Interpolation is used to express variable values at the control volume
surface in terms of the center values and suitable quadrature formulae are applied
to approximate the surface and volume integrals. An algebraic equation for each
of the control volumes can be obtained, in which a number of the neighboring
nodal values appear.
As with other discretization methods, a numerical grid must be initially defined
to discretize the physical flow domain of interest. For the finite-volume method,
there is the possibility of representing the grid by either structured or unstructured
mesh.
For illustration purposes of the finite-volume method, we consider a typical
representation of structured (quadrilateral) and unstructured (triangle) finite-
volume elements in two-dimensional shown in Figure 4.11 for the discretization of
the partial differential equations.
In particular, the so called Viscous hybrid meshes have proved to be suitable
for external aerodynamic purposes. This kind of mesh features several layers of
prismatic elements along walls, with tetrahedral elements in the core flow region.
4.5 The Finite Volume method: an overview 77

Figure 4.11: A representation of structured and unstructured mesh for the finite-volume
method (full symbols denote element vertices and open symbols at the center of the control
volumes denote computational nodes).

Compared to all-tetrahedral meshes, viscous hybrid meshes result in dramatic


savings, with far fewer elements required to accurately resolve boundary layers and
give good near-wall prediction of shear stress, heat transfer, and flow separation.
Let us consider the two unstructured grid depicted in Figure 4.12. In the case
of the fully tetrahedral unstructured grid (Figure 4.12, on the left), the demand of
small y+ (please read notes included in [7] and [8] for a deeper insight) would
translate into an unacceptable grid distortion near the wall. On the contrary, the
hybrid mesh depicted in Figure 4.12, on the right, allows to capture the near-wall
fluid dynamic phenomena by means of the thin prismatic layers, without the
need to increase the grid density in the core region. As a result, boundary layer
is captured in an effective way, while containing at a minimum the number of
elements.

Let us consider the two-dimensional continuity equation for an incompressible


flow:
∂u ∂v
+ = 0. (4.55)
∂x ∂y

Integrating it on a generic control volume yields the following expression,


78 4. Introduction to CFD

Non- structured grid: Non- structured grid:

Fully thetraedral Hybrid grid

Figure 4.12: A comparison between a fully tetrahedral unstructured grid (ON THE LEFT)
and a viscous hybrid mesh (ON THE RIGHT).

which is applicable to both structured and unstructured grids:


Z Z
∂u ∂v
dV + dV = 0. (4.56)
V ∂x V ∂y

where V stands for Volume.

Now, the above equation can be discretized by applying Gauss’ divergence


theorem to the volume integral (where A stands for Area, and N means the total
number of cells),
Z Z N
∂u
∑ ui Ai
y
dV = udA x ≈ (4.57)
V ∂x A k =1
Z Z N
∂v
∑ vi Ai
y
dV = vdAy ≈ (4.58)
V ∂y A k =1

Finally, the following equation is obtained:


N N
∑ ui Ai ∑ vi Ai
y y
+ =0 (4.59)
k =1 k =1

Eq. 4.59 represents the algebrized or discretized form of the continuity equation.
Now let us consider an elemental control volume of the two-dimensional structured
grid shown in Figure 4.13.
The centroid of the control volume is indicated by the point P, which is sur-
rounded by the adjacent control volumes having their respective centroids indicated
4.5 The Finite Volume method: an overview 79

Figure 4.13: Control volume for the two-dimensional continuity equation problem. Image
taken from [92]

by the points: east, E; west, W; north, N; and south, S. The control volume face
between points P and E is denoted by the area Aex . Subsequently, the rest of the
control volume faces are Aw x , Ay , and Ay , respectively.
n s
The analysis starts by introducing the control volume integration, which forms
the key step of the finite-volume method. Applying Gauss’ divergence theorem
yields the following expressions,
 =0 =0 
4
1
∑ ui Aix = ∆V ue Aex − uw Awx + un Anx − us Asx 
z }| { z }| {
(4.60)
k =1

=0 =0
 
4
1  z }| { z }| {
∑ vi Ai = ∆V ve Ae − vw Aw +vn An − vs As 
y y y y y
(4.61)
k =1

For the structured uniform grid arrangement, the projection areas Anx and Asx in
y y
the x direction, and the projection areas, Ae and Aw in the y direction, are zero.
Since the grid has been considered to be uniform, the face velocities ue , uw , vn , and
vs are located midway between each of the control volume centroids, which allows
us to determine the face velocities from the values located at the centroids of the
control volumes.
By assuming a linear velocity profile,
u P +u E u P + uW u P +u N u P +uS
ue = 2 uw = 2 un = 2 us = 2

By substituting the above expressions to the discretized form of the velocity


first-order derivatives, the final form of the discretized continuity equation becomes
80 4. Introduction to CFD

uP + uE u P + uW uP + u N u P + uS
       
y y
Aex − x
Aw + An − As (4.62)
2 2 2 2
y y
From Figure 4.13, Aex = Aw
x = ∆y and A = A = ∆x, the above equation can
n s
then be expressed by

uP + uE u P + uW uP + u N u P + uS
       
∆y − ∆y + ∆x − ∆x (4.63)
2 2 2 2

and reduced to

u E + uW v N + vS
   
∆y + ∆x = 0 (4.64)
2∆x 2∆y

Eq. 4.64 constitutes the algebraic form of the continuity equation.


It is important to note that the above equation referred to the central node of the
control volume depicted in Figure 4.13, i.e. P, contains the unknowns concerning
the adjacent points, that is N, S, W, and E. Therefore, for a n cells grid, the finite
volume method yields a system of n similar equations, to be solved after imposing
the so called boundary conditions at the domain boundaries.
Chapter 5

A Survey of the Techniques for


Drag Reduction

As just clarified in the abstract, the aim of this Thesis is to build up an auto-
matic optimization procedure, applicable to aircraft and rotorcraft aerodynamic
components, and to prove the effectiveness of that procedure by performing
an optimization run on a real industrial product. To achieve the latter purpose,
AgustaWestland has chosen the ERICA tilt-rotor intake configuration as a test
case. Therefore, it seems natural, at this point, to provide some notions about the
tiltrotor concept.
Moreover, for completeness of informations, a detailed survey of the most
promising techniques suitable for External Aerodynamic drag reduction is pro-
vided. For conceptual clarity, the different devices will be classified based on
the kind of drag to be lowered. As far as their application is concerned, both
advantages and disadvantages will be discussed with critical sense. For a deeper
overview, reader is referred to the precious notes given in the Von Karman Instiute
Lecture Series “Flow Control: Fundamentals, Advances and Applications” [33].

5.1 The Tilt-Rotor Concept


The peculiar characteristic of the tilt-rotor is the capability to take-off and land
like an helicopter and at the same time to cruise like an airplane by tilting the rotor
nacelles perpendicular to the flight direction, in helicopter mode, or parallel to the
flight direction, in aircraft mode. Moreover, the new tilt-rotor concept, of which
ERICA (Enhanced Rotorcraft Innovative Concept Achievement) is a technological
demonstrator (Fig .5.1), has the additional capability to tilt the outboard portion of
the wings independently from the nacelles, depicted in Figure 5.3 [49, 37].
In Figure 5.2 [66], the ERICA configuration as a function of the nacelle tilt
angle is depicted. This feature leads in some benefits about the thrust loss due
to rotors down-wash on the wings in helicopter mode, giving the opportunity to
reduce the rotor diameter, so as to improve the cruise performance. In addition,
its independent tilt freedom allows the wings to avoid stall, then improving the

81
����� 82 5. A Survey of the Techniques for Drag Reduction

operational capability of the aircraft during the helicopter-airplane conversion


phase. Moreover, the smaller dimensions of the rotors allow a tilt-rotor like ERICA
to ����
perform STOL (Short Take-Off and Landing) ����
operations, that means that also
take-off and landing in airplane mode is achievable [37].

����

Figure 5.1: An artistic impression of the ERICA concept (taken from [37]).

CARATTERISTICHE DI PROGETTO INNOVATIVO

90o Helicopter Mode

����
Nacelle
Angle Conversion
� Mode

0o
Aeroplane Mode

CRUI
Figure 5.2: Different Viaggio
ERICA della
possible Ricerca in Italia
configurations (taken-from [66]).
11

Milano – 17 giugno 2005


�������� The most suitable mission profile, for a tilt-rotor with these characteristics, is
�����������
the point to point service operated from and to heliports, but also to and from short
��������
field in order to increase the range and/or payload capabilities. A comparison
of the estimated operative costs among the presently operating aircrafts and tilt-
���������
rotors shows a competitive role of tilt-rotor with respect to helicopters for all
those missions whose required range is grater than 150 nautical miles, while the
competition with airplanes must be played on the field of tilt-rotor operational

�����������
���
5.2 Power requirements 3.2 Introduction to Intake Aerodynamics
83

FigureFigure ERICA
3.3:tiltable
5.3: ERICA wingtiltable wing
(taken from [1].
[37]).

3.2 Introduction
flexibility to Intake
that allows the minimization of the groundAerodynamics
infrastructure needs.

FromPower
5.2 the propulsive system point of view, a tilt-rotor is equivalent to a
requirements
turboprop-trusted aircraft. The attractiveness of turbo-propeller engines as com-
pared Thewithpurpose of an engines
turbofan aircraft, such
liesasinthetheir
futurehigher
European civil tilt-rotor
by-pass ratio based on
and thereby the
ERICA architecture, is to transport people from place to place
higher propulsive efficiency consequent to the smaller exhaust velocities. in a short time. To
do this, it must take off from land, climb to a desired flying altitude, cruise at a
As all the other air-breathing engines, turbo-propellers must be supplied with
prescribed distance, manoeuvre, descend and finally land at the desired destination.
air An
from the atmosphere in which the aircraft is operating. This supply duty is
aircraft is considered to have merit insofar as it performs these things quickly,
thesafely
aim of
and,thebyair
no intake.
means the least important, with the minimum expenditure of
The intake aerodynamic
energy. The quantitative measure behavior
of the waycan be roughly
it performs understood
such function is known byasanalyzing
thetheinteraction
performance between theA simple
of the aircraft. intakeand andaccurate method
two flow to understand
streams from an it is undisturbed
the
energy method
condition [36]. This
far ahead frommethod is based on the fact that, in steady flight, a power
the aircraft:
balance exists for an aircraft. As far as cruise flight condition is concerned, the
lift Internal
1. force L counterbalances the airplane
flow: it enters weightand
the intake W, whereas thrust T forcetask
the aerodynamic has to
is to mini-
exactly overcome the total drag force D. Therefore, two equations must be fulfilled,
mize the total pressure loss and maximize the flow uniformity with which
in the case of steady flight condition:
it reaches the engine face. This properties are vital from the engine perfor-
mance and stability pointWof= view;
L T=D (5.1)

By External
2. multiplying the
flow:second member
it passes of Eq. 5.1
around thebyintake
the factor V (speed),
as part of theantotal
equation
flow over the
between power quantities is found, that is
whole aircraft. It is important by means of the intake-aircraft integration
and its influence on the aircraftVD =drag.
VT (5.2)

where
The VD isofcalled
subject required
intake power, whereas
aerodynamics is the available
VT study
is the of both the Ifinternal
power. steady and the
external flow as defined in this way, with the scope of quantified their influence
on the aircraft efficiency and performances.
The system requirements, for an aircraft intake design, depend on the aircraft
mission specification. In the follows the main design goals for intake aerodynamics
are summarized:
84 5. A Survey of the Techniques for Drag Reduction

level flight is to be maintained, these powers must be equal [62]. Eq. 5.2 clearly
shows that the power required for level flight directly depends on the drag force.
Hence, a reduction of the drag force will translate into lower power consumption.
Conventionally, the total power required by the aircraft during steady flight
is split into three contributes. These power-absorbing elements are, according to
Gessow’s terminology:
• Induced-drag power;

• Propeller blades profile-drag power;

• Parasite-drag power.
Induced-drag power is the fraction required to produce lift, and can be reduced
by optimizing the shape and geometry of the lifting parts; it is due to both rotorcraft
body and rotors.
Profile-drag power is the power required to drag the propeller blades through
the air, and can be lowered by means of a proper shaping of the blades.
Parasite-drag power is, generally speaking, the power required to move a solid
object through a fluid. As far as tilt-rotor is concerned, this power arises from the
necessity of move the airframe and the rotating non-lifting components through
the air.
Parasitic drag is made up of many components, such as pressure drag, skin friction
drag and interference drag, which will be introduced in the next Section. In addition,
assuming that both transmission losses and various auxiliary equipment losses
account for a fixed fraction of the required power PR , the total shaft power PS
required of the engines is
PR
PS = + PA (5.3)
ηT
where PA is the power required to operate accessories, and ηT is the transmis-
sion efficiency.
Moreover, installation of the engines on the airframe generally results in a per-
formance deterioration when compared to the engine manufacturer’s performance
specifications. Losses associated with the engine installation can be divided into
inlet losses, exhaust losses, and losses due to bleed air extraction (e.g. air extraction
from the compressor for anti-ice protection). Engine installation effects are not
to be forgotten, since they play a smaller but not negligible role in the overall
power expenditure (as shown in Table 5.1) and they will be evaluated and reported
in a separate document. As high speed typical of cruise condition are reached,
the role played by parasite drag in the overall power balance becomes more and
more fundamental. Hence, a great reduction in power requirement is expected
to be achieved by improving the aerodynamic efficiency of the tilt rotor fuselage.
In Table 5.1, an example of power required breakdown is provided, showing the
different contributions to the overall power balance.
Table 5.1 clearly shows that fuselage parasite drag and body-induced drag
provide by far the greatest contributions to the overall power balance. Hence, an
airframe drag reduction is expected to be useful to gain an actual power reduction.
5.2 Power requirements 85

% Required Power
Engine Installation 8%
Fuselage 37%
Rotors (induced, viscous, vortex) 15%
Transmission 3%
Accessories 7%
Body Induced 30%

Table 5.1: Power required breakdown.

5.2.1 Fuselage and body induced drag breakdown

After performing power-required breakdown, a drag breakdown is to be carried


out. To this purpose, fuselage and body induced drag are taken into account only,
in order to put in evidence the contribution of the fixed and non-lifting components.
Table 5.2 provides fuselage and rotating non-lifting components drag breakdown
on a Tilt-rotor (ERICA configuration). The values, even if obtained on a scale model,
are referred to the full scale size. The different contributions are presented in terms
of equivalent flat-plate area, instead of drag coefficient. The term “flat-plate” is
the area of a flat plate having a CD of 1.0 with the same drag of the shape under
consideration. In other words, it represents the ratio of the Drag over the dynamic
pressure and is a meaningful way to compare bodies flying at the same speed
and density. The reason why drag coefficients of the single components will not
be considered is that they cannot be added since reference areas are different.
Considering flat plate area f of each component is a convenient way of handling
the drag, since they can be added to give the total f of an airplane, which, in turn,
is proportional to the drag force.

Fuselage and body induced component Equivalent flat-plate area, f [m2 ] Contribution [%]
Base fuselage 0.627 19.5
Fuselage-wing fairings 0.128 4.1
Sponsons16 0.403 12.5
Wings 0.691 21.6
Nacelles and spinners 0.211 6.6
Rotor hub and stubs 0.941 29.3
Fin 0.122 3.8
Tailplane 0.090 2.7
TOTAL 3.213 100

Table 5.2: Fuselage and body-induced drag breakdown. (NICETRIP wind tunnel tests,
2008. [67])

16 Sponsons are projections from the sides of an aircraft or helicopter, for protection, stability, or the

mounting of equipment such as armaments or lifeboats, etc. They are often used in larger helicopters
where the internal space of the sponson can be used for fuel or to house landing gear without
reducing cargo or passenger space in the fuselage
86 5. A Survey of the Techniques for Drag Reduction

As can be clearly seen from Table 5.2, rotor hub and stubs, wings and base
fuselage represent by far the greatest contributions to parasite and body-induced
drag. The next Section will provide an overview of the several drag reduction
techniques employable mainly on non-rotating components of a next-generation
Tilt-Rotor.

5.3 Overview of aircraft drag sources


Although the relative importance of different drag sources varies for each
Tilt-rotor
aircraft type and mission that is flown, fuselagebreakdown
a representative drag reduction
is shown in TLRG0000K019
-
Figure 5.4. The most important contributions to the total drag are the following:
Technology Review in support of Rev. A
1. Skin friction drag due to viscous boundary
Cleanlayer
Skyformation
projectand development;
Pag. 74
2. Pressure drag due to open separations in the afterbody and other regions; of 76
3. Interference effects between aerodynamic components;

4. Lift induced drag due to the conserved circulation developed around the
wings;

5. Miscellaneous effects such as roughness effects and leakage, etc;

6. Wave drag due to compressibility effects at near-sonic flight conditions;

Wave
Roughness
Skin friction
Miscellaneous
Interference
Afterbody

Lift-induced
drag

Figure 5.4: Contributions of different drag sources for a typical transport aircraft (taken
from [1]).

The greatest contribution arises from turbulent skin friction drag, a fact that
has justified the impetus for most of the friction drag reduction work that will be
described here. The next most significant contribution comes from the lift induced
drag and this, together with the friction drag, accounts for about 85% of the total
aircraft drag. Interference drag, wave drag, and miscellaneous effects account for
5.4 Skin friction drag reduction 87

the remainder. In particular, wave drag will not be taken into account since it plays
a negligible role in the overall drag of a tilt-rotor fuselage.

5.4 Skin friction drag reduction


Skin friction drag is the component resulting from viscous shear stresses, whose
integral is resolved in the drag direction. For the reduction of this kind of drag,
either (both, in the sense that will be clarified soon) of two different philosophies
may be followed:

1. The first consists in stabilizing the laminar boundary layer (BL) as to prevent
or delay transition to a turbulent one. In fact, because of the greater velocity
gradient characterizing turbulent layer than in a laminar one (due to the
more agitated motion in a turbulent flow), the frictional effects are more
severe for a turbulent BL;

2. An alternative philosophy for friction reduction that has recently emerged


is to accept the inevitability of turbulent flow and to attempt to modify or
interact with the turbulent structures so as to reduce the friction.

The two precedent methods are not in conflict with each other, but are com-
plementary since the former are exploitable on the wings, where the boundary
layer, at least near the leading edge, keeps laminar, whereas the latter is applied
mainly on fuselage where turbulent boundary layer exists. As regards the first class
of methods, they can be divided into active and passive, depending on whether
additional power from the propulsion units is required or not.

5.4.1 Laminar Flow Control (LFC)


An interesting active method for preventing laminar to turbulent transition is
the so called Laminar Flow Control (LFC), which means the maintenance of laminar
flow through the use of wall suction. Laminar to turbulent transition is a complex
phenomenon, which relates to instabilities and contaminations encountered by the
flow along the walls. Instabilities and contaminations can be classified as follows
([99, 75]):

• Streamwise instability;

• Cross flow instabilities;

• Leading edge contamination.

Streamwise instability is caused by amplification of the Tollmien-Schlichting waves


and is responsible for transition when the sweep angle φ is less than 25°. The point
of transition depends on the flow Reynolds number, pressure distribution and the
presence of surface roughness elements.
Cross-flow instability has its origin in high cross flow (φ ≥ 25°), which makes
high level Reynolds numbers the first cause of transition. This instability develops
88 5. A Survey of the Techniques for Drag Reduction

around the leading edge, where acceleration is high. It is a major concern for high
Reynolds number flow in subsonic transport aircraft, such as a Tilt-rotor.
Leading edge contamination (also calledfuselage
Tilt-rotor attachmentdrag contamination) is the third
line reduction
TLRG0000K019
-
cause of transition, and by no means the least. Attachment line contamination is
the phenomenon by which turbulentTechnology Review
air at the in support
wing root (comingoffrom a turbulent Rev. A
Clean Sky project
boundary layer on the fuselage) is propagated along the attachment line of a swept
Pag. 74
leading edge, causing the flow over the whole of the wing (or empennage surface)
of 76
to become turbulent. In the case of an engine nacelle, which has no leading edge
sweep, this kind of transition mechanism is not present.
These three sources of transition, which are frequently found together and
prematurely trigger transition near the leading edge, are well documented in the
open literature ([85, 47, 99]) and are depicted in Figure 5.5.

Figure 5.5: Transition mechanism on swept wings (taken from [85]).

Suction may be implemented in the form of distributed porosity over the


surface, in the form of a series of spanwise-running slots or in the form of discrete
drilled holes. Porous materials like foams, fabrics, nylons and sintered meshes,
tend however to have little structural integrity and shear stiffness, rendering them
unsuitable for use on their own as a skin material, unless supported by a load
bearing underlayer. Distributed suction acts in two main ways to suppress laminar-
turbulent transition [41]: first, it reduces the boundary-layer thickness; second,
it creates a much fuller velocity profile within the boundary layer, somewhat
similar to the effect of a favourable pressure gradient. This makes the boundary
layer much more stable with respect to the growth of small disturbances (e.g.
Tollmien-Schlichting waves, cross-flow instabilities, etc.). In Figure 5.6, the LFC
working principle is depicted. Although suction modifies the velocity profile in the
boundary layer, in such a way that in some cases the local skin friction coefficient is
reduced by more than 10% [75], several reasons make this technique unattractive,
that are itemized in the following:
• It requires a very fine definition of the suction system parameters. These
parameters involve hole size, suction flow rate, hole spacing and geometry,
and hole inclination [48];
Questo documento contiene dati ed informazioni di proprietà Agusta che ne vieta la divulgazione e
la riproduzione parziale o totale senza la sua autorizzazione scritta.
This document contains data and information which are the property of Agusta. Agusta forbids its circulation and
reproduction, whether in part or in its entirety, without a written authorization from the company.

AG-TEC168 Rev. 07
5.4 Skin friction drag reduction 89

• The results obtained depend very much on the state of the surface and on
the absence of contaminations (e.g. insects or dust), which is a not negligible
issue when considering its application to Tilt-rotor wings;

• It is characterized by high maintenance costs related to the high amount of


energy necessary for the suction;

• Acoustic waves and vibrations on the wings and tail assembly could impair
the effectiveness of LFC system, since laminar to turbulent transitions would
be promoted [48]. This is an issue for Tilt-rotor aircrafts, since tiltable rotors
are mounted on the wings;

• Contamination due to fuselage boundary layer is a real concern;

• From an aerodynamic viewpoint, the greatest difficulty lies in being able to


confidently predict where transition will occur, although CFD techniques
could be helpful with regard to this.

An essential problem with any laminar flow condition is its susceptibility to


dirt and other particulates, Tilt-rotor
such as fuselage
insect debris accumulatingTLRG0000K019
drag reduction near the leading
-
edge during low altitude flight. These can trip the flow to turbulence, which will
Technology Review in support of
then spread over a wide area of Cleanthe wing. To avoid this, close manufacturing
Sky project
Rev. A

tolerances must be followed and some kind of in flight cleaning system or Pag.leading-
74
of 76
edge protection must be employed. Moreover, de-icer inserts on shields for ice
protection and supplementary nozzles for protection from insects and ice are
required.

Figure 5.6: LFC system designed by McDonnell- Co. (taken from [48]).

In Figure 5.6, the Douglas concept for LFC application is showed; it involves
an electron-beam-perforated titanium sheet bonded to a fiberglass-corrugated
substructure. Suction was applied from just below the attachment line back to
the front spar. A Krueger shield was used at the leading edge to deflect or block
insects. TKS anti-ice system was used on the Krueger shield, and a spray nozzle
system was appended to the back of the Krueger shield as a backup system for
90 5. A Survey of the Techniques for Drag Reduction

anti-insect and anti-ice protection of the leading edge. The Krueger shield was
typically retracted after reaching an altitude of 6000 ft, with the goal of leaving an
insect-free leading edge for cruise flight. This LFC system was employed during
the Jetstar LFC Leading-Edge Flight test Program (1983-1986) [41] and showed the
effectiveness of LFC systems for airline service.

5.4.2 Natural Laminar Flow (NLF)


Among passive techniques exploitable to BL control, attention has to be paid
on Natural Laminar Flow (NLF), a technique that uses optimized pressure gradients
to maintain laminar flow over a large part of the wings. To obtain these results,
maximum profile thickness must be as far aft as possible in order to create extensive
regions of favourable pressure gradient over the wing surface. Although this
passive technique is less expensive than LFC one, it presents some limits in term
of sweep angle and Mach number. In fact, if the Reynolds number exceeds a
critical value, Tollmien-Schlichting waves within a laminar BL will be amplified
as they propagate downstream, resulting in transition. Moreover, attachment
line contamination is the phenomenon by which turbulent air at the wing root
(coming from a turbulent boundary layer on the fuselage) is propagated along
the attachment line of a swept leading edge, causing the flow over the whole of
the wing (or tail assembly surface) to become turbulent. These two phenomena
occur when Mach number and sweep angle increase excessively. For these reasons,
passive laminar flow control is suitable for regional aircraft only [70].
Even though the concept can be employed without the need for considering the
attendant weight and structural penalties associated with the LFC suction system,
some disadvantages characterize this device, that are:

• Structural weight penalties to be taken into account;

• Permissible wing sweep is limited by the onset of cross-flow instability at the


leading edge;

• The challenge is to design wings maintaining natural laminarity with a


high performance level (as, for example, the Mach number). The Mach
number effect frequently increases the sweep angle which is contrary to the
application of this technology;

• Only true savings, including the effect of the system for cleaning the leading
edge required to prevent insect contamination, ice formation, and the pro-
duction costs of such wings will produce tangible benefits where the Direct
Operating Costs is concerned;

• The high production costs of such a wing system. For example, to have a
laminar upper surface it must not have any roughness greater than 0.1 mm
[75];

• Contamination due to fuselage boundary layer is a real concern;


5.4 Skin friction drag reduction 91

• The presence of the wing tip-mounted propellers impairs the effectiveness


of NFL
P1: MBL/ary devices, QC:
P2: MBL/plb since an early
MBL/abe laminar to turbulent transition is promoted
T1: MBL
[44, 18].
November 21, 1997 16:48 Annual Reviews AR049-01

5.4.3 Hybrid Laminar Flow Control (HLFC)


An attractive
2 method for preventing laminar to turbulent transition is called
JOSLIN
Hybrid laminar flow control (HLFC), an ingenious compromise between Natural and
Controlled 1.1 What IsThis
laminarity. Laminar Flow Control?
compromise makes it possible to do away with the
What is LFC? LFC is an active boundary-layer flow control technique (usu-
disadvantages of the two technologies described above. By means of suction in
ally steady suction) employed to maintain the laminar flow (LF) state at chord
the area between
Reynolds the first beyond
numbers 15-20% of that
those theare
chord andcharacterized
normally a favourable pressure
as being tran- gradient
as for NLF to about 50% of the chord, the effects of cross flow instabilities are
sitional or turbulent in the absence of control. Understanding this definition is
minimized anand important first step toward understanding the goals of the technology. Often,
a natural laminar flow may then develop in the mid-part of the
one mistakenly assumes that LFC implies the relaminarization of a turbulent
wing profileflow.
[99]. Thearekey
These two features of physics
different flow HLFCphenomena,
are [48]: and although the same
control system may be employed for both problems, the energy requirements
Annu. Rev. Fluid Mech. 1998.30:1-29. Downloaded from arjournals.annualreviews.org

• Suction
foris required only
relaminarization couldin the leading-edge
typically region,greater
be an order of magnitude in order to suppress the
than those
required for LFC. Finally, LFC is a capability
cross-flow vortices and leading-edge contamination; that is designed to benefit the
by Universita degli Studi di Padova on 09/02/08. For personal use only.

aircraft during cruise.


A significant advancement
• NLF is maintained over the made
winginthrough
the development
properof LFC is the technol-
tailoring of the geometry
ogy of hybrid laminar flow control (HLFC). HLFC integrates natural laminar
(effectflow
on (NLF)
free-stream flow pressure);
with LFC to reduce suction requirements and system complexity
(Figure 1). NLF employs favorable pressure gradients to delay the transition
• The HLFCprocess,wing design
is sweep limited,has
andgood
usuallyperformance in cruise
has poor off-design the turbulent
aerodynamicmode.
performance. LFC is complex, involving suction (and ducts, flutes, and pump
In Figure 5.7, a conceptual
source) over the whole comparison
wing chord (oramong three previous
engine nacelle). techniques
The key features of is pre-
sented. HLFC are (a) suction is required only in the leading-edge region ahead of

Figure 1 Schematic of NLF, LFC, and HLFC concepts for wing (Collier 1993). Diagrams show
suction locations and surface pressure coefficient (C p ) versus chordwise extent (x/c).
Figure 5.7: NLF, LFC and HLFC concepts for wing (taken from [48]).

Using HLFC technology, the region of favourable pressure gradient extends


until 50% of chord length. The advantage of using this device consists in that
it is a HLFC system that avoids some of the problems associated with LFC and
NLF. Suction is applied only at the leading edge to minimize cross-flow instability.
Control of the instabilities in the mid-chord region is achieved with tailoring of the
92 5. A Survey of the Techniques for Drag Reduction

pressure gradient as with NFL. In this way, a larger wing sweep can be achieved
for transonic flight than with NLF, and the weight penalties are not as great as
for LFC. As far as its disadvantages is concerned, it has to be underlined that
the effectiveness of this system can be impaired by cloudy conditions or rainy
weather (because of the lost of laminarity during these conditions). Moreover,
contamination due to fuselage boundary layer can be a real concern. Finally,
the maintainability and reliability of suction surfaces and the optimization of
suction rate and distribution remain the main problems regarding this preventing-
transition method.

5.4.4 Riblets
An alternative approach toTilt-rotor
the reduction of skin
fuselage dragfriction is based not upon
reduction
TLRG0000K019
trying to maintain laminar flow, but on attempting - to modify the turbulence in
some way so as to reduce friction. Possible
Technology approaches
Review may of
in support be passive, as in Rev. A
Clean Sky
the case of the riblets and large eddy break-up project
devices (LEBUs) etc., or active as in
the case of the synthetic boundary layer. Over the years, extensive research
of 76
onPag. 74
riblets has been carried out at the NASA Langley Research Centre (USA) and
ONERA/CERT (France). Riblets, (micro-grooves on the surface, aligned to the free
stream direction, see Figure 5.8), have been studied deeply ([70, 95, 96]) and the
results from these studies have been sufficiently promising and encouraging that
the concept has been evaluated in flight tests. Despite worldwide research during

Figure 5.8: Sketch of riblet geometry (taken from [95]).

the last 15 years, detailed mechanisms by which riblets reduce the wall shear
stress are not clearly understood. Several mechanisms have been suggested which
include: weakening of the bursting process near the wall, significant retardation
of the flow in the groove valley dominated by viscous effects, an increase in the
sub layer thickness inhibition or restriction of spanwise motion of longitudinal
vortices [70]. It is likely that many of the above flow features have their subtle role
in altering the wall shear stress. It seems that highly turbulent activity of boundary
layer is “lowered” in such a way that skin friction is reduced. Apart from these
fluid-dynamics considerations, the aerodynamic advantage procured by riblets
has been clearly demonstrated during many wind-tunnel and flight tests. It is
worth noting that, despite the increase in the wetted area, a net drag reduction
5.4 Skin friction drag reduction 93

occurs when riblets technique is exploited. This is the reason why an optimization
procedure has to be employed, when designing the correct riblets shape.
The main advantages of their application are:
• Their effectiveness is not affected by pressure gradient [70];

• Their effectiveness is not affected by compressibility;

• Their effectiveness is not affected by moderate misalignment to the mean


flow [70];

• They are easy to install: plastic riblet films with symmetric V grooves manu-
factured by the 3M© are suitable to this purpose.
On the other hand , the disadvantages are
• The microscopic structures of riblets are highly sensitive to dirt and mechani-
cal degradation;

• Acceptable installation and removal times must be achieved. Airbus In-


dustry© and 3M© has defined a technique that has cut down 3M’s first
assessment by a factor 10. Concerning removal, aqua stripping seems the
most suitable technique [75];

• Their effectiveness is affected by high misalignment to the mean flow [70]. At


25°-30° their capacity for drag reduction is significantly reduced; moreover,
with misalignment, a mean pressure difference across each riblet occurs; such
a pressure difference can result in a significant (because of the large area
involved) force normal to the line of the riblets, which, in turn, can contribute
directly to the drag.

5.4.5 LEBUs
Another passive techniques aiming at modifying turbulent boundary layer
structure in such way that there is a net drag reduction, is represented by Large
Eddy Break Up (LEBU) devices. They are designed to alter, sever, or break-up
the large eddies that form in the outer regions of a turbulent boundary layer.
A typical arrangement consists of one or more splitter plates placed in tandem
in the outer part of a turbulent boundary layer, as sketched in Figure 5.9. An
analytical attempt to explain the mechanisms involved with the LEBUs is noted
here. Essentially, the LEBU acts as an airfoil on a gusty atmosphere, and a vortex
unwinding mechanism is activated. The vorticity shed from the LEBU’s trailing
edge because of an incident line vortex convected past the device tends to cancel
the effect of the incoming vortex and to reduce the velocity fluctuations near the
wall. The main advantage characterizing LEBU is that their effectiveness seems not
to be affected by adverse pressure gradient [35]. However, they are not effective at
Reynolds and Mach numbers typical of flight conditions [35]. Moreover, what is
difficult is to ensure that the device own skin-friction and pressure drag do not
exceed the savings.
Pag. 74
of 76

94 5. A Survey of the Techniques for Drag Reduction

Figure 5.9: Sketch of a tandem arrangement of a LEBU device (taken from [35]).

5.5 Pressure drag reduction


Pressure drag (also named shape drag) is the drag of a body resulting from the
integrated effect of the static pressure acting normal to its surface resolved in the
drag direction. In an aircraft, two typical shape drag effects decrease the flight
efficiency, i.e. flow separation in the aft region of the fuselage, and near the trailing
edge of the wings, both occurring during off-design operating conditions (e.g.
during take-off phases). As far as afterbody drag is concerned, it is a small but
significant contribution to the overall aircraft drag, as depicted Figure 5.4. This
kind of drag is related to the presence of the upswept regions, essential in order
to meet operational requirements and take-off rotation, and anyway present in
a tilt-rotor. The basic features of the flow field typical of upswept fuselages are
also shown in Figure 5.10 [48], for a typical aircraft. It is characterized by a 3-D
boundary layer with significant cross flow regions on the fuselage. This boundary
layer separates into a pair of counter rotating-vortices trailing downstream. The
Questo documento contiene dati ed informazioni di proprietà Agusta che ne vieta la divulgazione e
flow is analogous to the
la riproduzione flow
parziale about
o totale senza la a sua
missile at high
autorizzazione scritta.angle of attack or the flow over
This document contains data and information which are the property of Agusta. Agusta forbids its circulation and
a delta wing,reproduction,
although in the present case a hard separation line does not exist. The
whether in part or in its entirety, without a written authorization from the company.

AG-TEC168 Rev. 07
total drag associated with this kind of flow can be split into two components. First,
the pressure drag arises because of the reduced pressures on the lower surface of
the fuselage. In addition, there is a considerable loss of flow energy in the form
of rotational kinetic energy of the vortex structures and this is manifested as a
vortex drag component. Depending on the geometry of the aircraft, the relative
contributions of each may vary.
In the following, a description of the main flow and boundary layer control
devices, suitable for either fixed wing or rotary-wings aircraft will be provided.

5.5.1 Synthetic Jet actuators


Among various flow control methods, synthetic jet actuation is one of the most
promising as it demonstrates a great potential for active flow control. Synthetic jets
are attracting increasing interest in the aerospace community because they have
many practical applications, including jet vectoring, control of separation, enhanced
5.5 Pressure drag reduction 95

Figure 5.10: The vortex wake behind an upswept afterbody (taken from [89]).

mixing, reduction of wall skin friction, and virtual aero shaping. In particular,
as far as separation is concerned, there is a strong motivation to manipulate or
delay its occurrence, since it means preventing large energy losses and in most
applications lift loss and drag increase.
As a practical tool, they are generally more attractive than steady blowing
or suction (e.g. LFC) because they do not require any complex fluid ducting or
plumbing system. Furthermore, synthetic jets have been found to achieve similar
effectiveness to steady blowing or suction with considerably smaller momentum
(and hence lower energy inputs) [81].
The synthetic jet is a zero net mass flux device, which consists of an oscillating
diaphragm inside a sealed cavity with a small orifice or slit through which an
air jet enters (suctioning) and exits (blowing) due to the diaphragm oscillatory
displacement. The expelled fluid forms a shear layer with the surrounding fluid
that results in a series of rolling vortices. If the oscillation of the diaphragm is
sufficient in amplitude and frequency, the vortices have sufficient inertia to escape
re-entrainment into the cavity resulting in an air jet consisting of propagating
vortices [6, 63].
Even though the net mass change of fluid through the cavity during each cycle
of the diaphragm motion is zero, the net momentum transferred into the fluid is
non-zero.
Zero net mass flux synthetic jet actuators produce an air jet with unique effects,
which are impossible to be obtained with steady suction or blowing. Figure 5.11
shows a schematic of a typical
Questo documento contiene synthetic jet or “zero
dati ed informazioni di net massAgusta
proprietà flux”chedevice.
ne vieta la divulgazione e
la riproduzione parziale o totale senza la sua autorizzazione scritta.
The vortical structures promote boundary layer mixing and hence momentum
This document contains data and information which are the property of Agusta. Agusta forbids its circulation and
exchange between the outer and inner parts of the boundary layer. The enhanced
reproduction, whether in part or in its entirety, without a written authorization from the company.

mixing AG-TEC168
causes theRev. reattachment
07 of the separated shear layer. Hence, the actuators
act as turbulators, energizing the boundary layer and thus eliminating massive
separation.
Synthetic jet systems have been produced using speakers, compressed air, air
pumps, and bimorph diaphragms. All of these techniques add weight, require real
estate, and add complexity to an aeroplane, making these options difficult to be
used in practice.
96 5. A Survey of the Techniques for Drag Reduction

Figure 5.11: Schematic of a synthetic jet (taken from [6]).

Piezoelectric actuators are instead an attractive solution because of their


lightweight and fast time response; they have no contacts, they are highly re-
liable, and most importantly they are cheap to produce [5]. When subject to an
electric field, piezoelectric actuators such as lead zirconate titanate ceramics pro-
duce mechanical strain or alternately generate an electric charge when subjected
to a mechanical strain. This property gives piezoelectric materials the ability to act
as actuators, hence to be employed in active flow control.
Synthetic jet systems have been tested on the upper surface of a NACA 0015
airfoil [98]. The synthetic jet actuation not only stabilizes the boundary layer
either by adding/removing the momentum to/from the boundary layer, but also
enhances mixing between inner and outer parts of the boundary layer. Hence,
downstream flow separation is prevented or postponed.
Main features of the flow over uncontrolled and controlled airfoils are revealed
in Figure 5.12, showing iso-surfaces of the instantaneous vorticity magnitude
overlapped with pressure contours predicted by the present LES. The vortical
structures present over the suction surface qualitatively indicate the degree of flow
separation. In the uncontrolled case (Figure 5.12a), flow massively separates from
the half aft portion of the suction surface while the flow separation is dramatically
prevented with the synthetic jet actuation in the controlled case (Figure 5.12b). The
present synthetic jet actuation
Questo documentowith the momentum
contiene coefficient
dati ed informazioni of 1.23%
di proprietà produces
Agusta che ne vieta la divulgazione e
la riproduzione parziale o totale senza la sua autorizzazione
more than a 70% increase in the lift coefficient. The drag coefficient is found to scritta.
This document contains data and information which are the property of Agusta. Agusta forbids its circulation and
decrease approximately by in15-18%
reproduction, whether part or inwith the synthetic
its entirety, jet actuation.
without a written authorization from the company.

AG-TEC168 Rev. 07
5.5 Pressure drag reduction 97

Figure 5.12: Iso surfaces of the instantaneous vorticity magnitude. (a) Uncontrolled case;
(b) controlled case. (taken from [98])

The effectiveness of synthetic jets for boundary layer flow control under an
adverse pressure gradient condition was investigated as well, by means of experi-
mental tests presented in [45]. The enhancement of this effectiveness by the T-S
instability waves was noticed and analyzed. The conclusions are summarized as
follows [93].

• The effectiveness of synthetic jets on boundary layer flow control can be


enhanced by the natural instability of the boundary layer flow;

• The synthetic jets are effective when the forcing frequency is low, in the range
of T-S frequencies;

• The effectiveness of synthetic jets may depend more on the forcing frequency
rather than on the forcing voltage;

• SyntheticQuesto
jets can play contiene
documento dual roles dati edin resistingdi separation
informazioni proprietà Agusta bycheaccelerating
ne vieta la divulgazione e
la riproduzione parziale o totale senza la sua autorizzazione scritta.
turbulence Thisand reducing
document contains turbulence
data and information in awhich
naturally turbulent
are the property boundary
of Agusta. layer;
Agusta forbids
reproduction, whether in part or in its entirety, without a written authorization from the company.
its circulation and

AG-TEC168 Rev. 07
• The effectiveness of the synthetic jet actuator in boundary layer control may
not be predictable with only the output of the synthetic jet actuator operating
in a condition without cross flow;

• Synthetic jet actuators operating at lower forcing frequency may be more


preferable than those at higher forcing frequency, because the audible noise
generated at higher frequencies is avoided.
98 5. A Survey of the Techniques for Drag Reduction

Synthetic jet actuators have shown their effectiveness in reducing pressure drag
from turbulent boundary layer separation also during wind tunnel tests and
numerical simulations performed on a helicopter fuselage model [40]. A similar
application on a tilt rotor seems to be plausible.
The main advantages connected with these devices are here reported:

• As an active flow control device, synthetic jet actuators allow to attain a large
effect using a small, localized energy;

• Since zero net mass flux is involved, the synthetic jet device has the advantage
of eliminating the need for plumbing connections and an internal air supply,
then allowing a reduction in weight and costs;

• Unlike Passive control devices, e.g. vortex generators, the synthetic jet device
does not introduce a drag penalty when the flow does not separate;

• Piezoelectric devices have the advantages of faster response, good reliability,


low cost, and a reduction in weight and space;

• Piezoelectric devices are characterized by an efficiency larger than 90% [93];

• Micro-electro-mechanical systems (MEMS) technology will allow producing


arrays of a large number of these actuators (necessary for turbulence control)
at a reasonable cost; item If the piezoelectric actuators were driven at a
frequency of operation (but out of phase) to counter a known noise source,
the technology certainly could be a viable noise reduction technology (active
noise suppression).

On the other hand, the disadvantages the designer has to deal with, are

• Designing such a system is difficult. The cavity volume, neck length, slot
size, diaphragm area, and frequency must be correctly designed, in order to
obtain good results;

• The complex unsteady flow is hard to be computed by means of traditional


computational fluid dynamics (CFD) techniques, and instead requires the un-
steady CFD codes employment, which, in turn, requires huge computational
resources and long turnaround time;

• Significant noise levels may be introduced by operating these devices at a


resonant condition.

5.5.2 Vortex Generators


As an example of flow control on afterbodies, one can consider the effect of
vortex generators on the flow along the aft fuselage of a transport aircraft and
the corresponding influence on drag. In general, the aft ends of transport aircraft
fuselages are tapered asymmetrically with pronounced upsweep of the lower
contour to facilitate rotating in the pitch lane on landing and take-off and to meet
Tilt-rotor fuselage drag reduction
TLRG0000K01
-
5.5 Pressure drag reduction
Technology Review in support of 99 Rev.
Clean Sky project
Pag.
operational requirements. The longitudinal pressure gradients due to the upsweep of 76
result in a rapid boundary layer growth and eventually in separation and vortex
shedding accompanied by an increase in fuselage form drag. The use of vortex
generators located, for instance, at the beginning of the upswept region as shown
in Figure 5.13, is promising to reduce pressure drag [85]. The basic principle which

Figure 5.13: Model mounted in the wind tunnel: detail of vortex generators (taken from
[85]).

suggests the use of vortex generators to energize a boundary layer in the presence
of an adverse pressure gradient is based on the enhanced mixing produced by the
vortices between the high energy flow at the edge of the boundary layer and the
low energy flow in the boundary layer. In this manner, the energized boundary
layer flow can withstand a higher adverse pressure gradient without the occurrence
of separation, which considerably increases the pressure drag. Vortex generators
have been tested also on a C-130 aircraft model to postpone separation [17].
The main advantages are:

• As a passive technique, they would require little maintenance;

• They do not require the expenditure of additional power from the propulsion
units.

The principal disadvantages are:

• They can introduce a drag penalty when the flow does not separate, e.g.
during cruise condition;

• Passive solutions like Vortex Generators cannot control the flow velocity for
different flight conditions because they are mounted to the surface perma-
nently and there is no possibility to change their shape or position during
the flight conditions.

In Figure 5.14, Vortex Generators of various shapes and sizes are depicted.

Questo documento contiene dati ed informazioni di proprietà Agusta che ne vieta la divulgazione e
la riproduzione parziale o totale senza la sua autorizzazione scritta.
This document contains data and information which are the property of Agusta. Agusta forbids its circulation and
reproduction, whether in part or in its entirety, without a written authorization from the company.

AG-TEC168 Rev. 07
100 5. A Survey of the Techniques for Drag Reduction

Figure 5.14: Different shapes of vortex generators (taken from [42]).

5.5.3 Active Vortex Generators


Active vortex generators (VGs) may be applicable for separation control during
aircraft take-off and landing, and drag reduction during aircraft cruise conditions.
Active VGs have a similar effect on the flow structure to passive VGs but these
devices potentially have an advantage over conventional VGs because they can
eliminate the parasitic drag that arises with VGs, and because the unsteady fluidic
actuator can induce slip velocities (which may reduce drag) during cruise condi-
tions. Pulsed, air-jet vortex generators (Figure 5.15) are placed upstream of the
region of expected flow separation. When operating, the vortex generators create
streamwise vortices in the lower/middle regions of the boundary layer which
have been demonstrated to lead to substantial improvements in the shape factor,
and hence reduced susceptibility to flow separation. Each actuator comprises an
orifice which is pitched at an angle 45° to the surface and skewed at an angle of
approximately 90° to the local onset flow. Typically the orifice has a diameter of
between approximately 5 and 15% of the local boundary-layer thickness. Each
orifice is supplied with a pressurized air mass flow, which is controlled by a valve
located below the orifice. Each orifice is therefore individually controllable and
capable of generating an air-jet on demand in response to a command signal.
Typically the actuator is required to operate at a frequency of up to 500 [Hz] with
duty cycles (ratio of the time on to the time off) of 50% or less. Each actuator
has to beQuesto
capable of generating
documento contiene dati an exit jet velocity
ed informazioni which is
di proprietà of the
Agusta chesame order
ne vieta of
la divulgazione e
la riproduzione parziale o totale senza la sua autorizzazione scritta.
magnitude Thisas the local
document freestream
contains velocity.which
data and information Foraretypical full-scale
the property application
of Agusta. Agusta forbidsinitsthe
circulation and
reproduction, whether in part or in its entirety, without a written authorization from the company.
region of a leading-edge slat or trailing-edge flap on a medium-sized transport
AG-TEC168
aircraft, Rev. 07orifice diameter would be of the order of 200-400 [µm] and
the required
the peak jet velocity would be of the order of 200-300 [m/s] [97].
The main advantage is:

• Compared with passive vortex generators, they can eliminate the parasitic
5.5 Pressure drag reduction 101

Figure 5.15: Schematic of the proposed separation control actuation concept (taken from
[97]).

drag that arises with VGs during cruise conditions.

The principal disadvantage is:

• Their effectiveness is closely related to the the aerodynamic optimization of


the flow-actuation mechanism [38].

5.5.4 Micro Vortex Generators


Unless large passive VGs which have been used on the aft portion of an aircraft
to improve overall performance of the aircraft, micro-VGs (less than one-third of
the boundary layer thickness) have been used on the wings of production aircraft
to prevent flow separation, primarily during take-off and landing conditions, and
to mitigate flap side-edge noise. Lin et al. [59] have shown that using micro-VGs on
the flap of a high lift system can mitigate flow separation, leading to a 10 percent
increase in lift and a 50 percent reduction in drag. The main advantage is:

• The size reduction significantly reduces the parasitic drag during cruise
conditions, and is enabled by the extreme fullness of the mean velocity
profile in a turbulent boundary layer.

The principal disadvantage is:

• Such devices must be placed closer to the nominal separation location and
therefore may documento
Questo be less suitable thandati
contiene larger devices for di
ed informazioni situations
proprietà in whichche
Agusta thene vieta la divulgazio
la riproduzione parziale o totale senza la sua autorizzazione scritta.
separation region is not relatively localized.
This document contains data and information which are the property of Agusta. Agusta forbids its circulation
reproduction, whether in part or in its entirety, without a written authorization from the company.

AG-TEC168 Rev. 07
102 5. A Survey of the Techniques for Drag Reduction

5.6 Interference drag reduction


When two shapes intersect or are placed in proximity, their pressure distribu-
tions and boundary layers can interact with each other, resulting in a net drag of
the combination that is higher than the sum of the separate drags. This increment
in the drag is known as interference drag.
This juncture drag is due to the occurrence of an unfavourable modification into
the local pressure field and the additional rapid straining of the highly dissipative
vorticity. The flow field in an unfilletted juncture region is characterized by the
formation of a horseshoe vortex structure ahead of the junction [82]. An example
of this type of flow occurs when the flow on the fuselage of an aircraft interacts
with the wing. The resulting vortex can also reduce the lift and increase the drag of
the aircraft. In addition, the regions of concentrated streamwise vorticity that trail
behind the juncture decay slowly and interact with components downstream, such
as tail assembly and engines. Therefore, the horseshoe vortex can also alter the
stability and control characteristics of the aircraft. Proper filleting and fairings can
reduce these effects. Another issue related to wing-fuselage or wing-empennage
junction, is the flow contamination over the wing or tail surfaces. This phenomenon,
by which the turbulence at the wing root is propagated along the attachment line
causing the whole flow to become turbulent, would impair the effectiveness of
drag reduction techniques, such as HLFC devices previously explained.
Actually, in order to obtain a region of laminar flow on a wing, it is first
necessary to ensure that the attachment line remains laminar, namely, to avoid the
problem of attachment line contamination.

5.6.1 Gaster bumps


A passive device suitable to avoid the attachment line contamination is the
so-called Gaster bump. This leading-edge device, a bracelet or a bump put on
the leading edge of the wing quite close to the fuselage, creates a stagnation
point on the attachment line, stopping the propagation of turbulence and allowing
a new laminar boundary layer to be generated downstream. For application to
real aircraft, the bump has also to be aerodynamically as smooth as possible to
minimize the additional drag due to its presence. Current CFD techniques are
useful tools in the designing of these anti-contamination devices [72].

5.6.2 Slot suction


An active method suitable to prevent turbulent boundary layer to spread
along the wing surface, thus preventing attachment line contamination, is the
localized wall suction. ONERA©carried out wind tunnel tests [72], which proved
the effectiveness of active attachment line anti-contamination devices (boundary
layer suction). However, it is necessary to undertake studies of roughness effects
and wing surface smoothness requirements even if it is clear that active devices
are much more robust towards surfaces defects.
5.7 Lift-induced drag reduction 103

Figure 5.16: Winglet attached to wing tip (taken from [47]).

5.7 Lift-induced drag reduction


The drag due to lift of a finite span wing is usually due primarily to the induced
drag associated with the shedding of vorticity along the span and, in particular,
at the tip. It is also due in part to an increase in the skin friction and form drag
associated with an increase in the lift coefficient (since the vortex structure is
concentrated at the tip and induces a downwash over the wing, which reduces
the effective angle of attack) [1]. To reduce the induced drag, wings of large aspect
ratio should be utilized since these enable the tip vortex structures to be separated
which reduces the strength of the average induced flow between them. However,
the aspect ratio employed in actual airplane configurations is usually constrained
by practical considerations, particular those associated with the wing structural
weight. When the aspect ratio is increased, the bending moments in the wing
structure also increase, with the resulting requirement for more structural material
in the wing. Since the performance of the airplane is determined by the structural
weight as well as the aerodynamic efficiency, the optimum aspect ratio for the best
overall performance must be a compromise [41].

5.7.1 Winglets
The addition of tip-mounted surfaces to a wing can reduce and diffuse the
vortex structures arising from the tips. Induced drag reductions result, but these
may be offset by unfavourable interference and viscous effects. The winglet is
one of the most known concept which can be thought as a device to increase
the effective span of the wing. As shown in Figure 5.16, the winglet is a small
wing mounted in the swirling flow at the wing tip. The lift on the winglet acts
as a side force and, with proper positioning of the winglet, it will have a thrust
component in Questo documento
the stream contiene
direction. Asdati
with edthe
informazioni
afterbodydi strakes,
proprietàthe
Agusta che ne vieta la divulgazio
structure
la riproduzione parziale o totale senza la sua autorizzazione scritta.
of the vorticesThisisdocument
somewhat diffused
contains dueinformation
data and to the winglets.
which are However,
the propertyone has toAgusta
of Agusta. be forbids its circulation
reproduction, whether in part or in its entirety, without a written authorization from the company.

AG-TEC168 Rev. 07
104 5. A Survey of the Techniques for Drag Reduction

careful during winglet design process, since such surfaces have significant effects
on the structural weight. Actually, loads on the vertical surfaces and the increased
loads on the outboard region of the wing associated with adding these surfaces
increase the bending moments imposed on the wing structure. Moreover, adding
winglets means an increase in the overall wetted area, hence an increase in the skin
friction drag. The main advantages are:
• A proper design of winglets can also improve tilt-rotor stability;

• No additional energy is required.


The disadvantages are:
• Adding winglets means an increase in the overall wetted area, hence an
increase in the skin friction drag;

• An increase in overall weight has to be taken into account.

5.8 Miscellaneous drag sources


Items included in the following discussion refer to the surface roughness and
leakage, protuberances, and momentum losses.
Surface roughness drag includes surfaces irregularities such as rivets heads,
seams, waviness in the skin, etc. It is reduced through the use of butt joints on all
exterior airframe structures, and flush rivets on streamlined components (fuselage,
tail assembly and nacelles). A study [34] indicates that the drag increment due
to the discrete roughness of streamlined V/STOL aircraft components is reduced
from 9 to 5% through the use of flush rivets over the entire aircraft. Protuberances
include such items as antennas, lights, handholds Pitot tubes and windshield
wipers. The related drag is relative large since they are bluff bodies. The use of
fairings and selective positioning can result in sizeable drag reduction benefits.
Leakage drag results from air that enters or exits the fuselage around access
doors, access panels and windows, which are to be sealed in order to reduce this
kind of drag.
Momentum losses occur when air stream is diverted into the aircraft for the
purposes of cooling the transmissions, hydraulic systems and engines, and pro-
viding air for heating-ventilations systems. The drag penalty can be minimized by
providing generous inlet radii to prevent air spillage from producing separation,
and direct the exhaust air from the system aft to recover a portion of the inlet
momentum less.
A momentum loss also results due to the change in velocity of the air entering
and exiting the engines at high speeds.

5.9 Regions of application


In Table 5.3, a resume of the tilt-rotor regions suitable for application is carried
out. Devices are classified depending on whether they are passive or active tech-
5.9 Regions of application 105

Device Energy requirements Region of appplication


LFC Active Wings, fin, tail-plane
NLF Passive Wings, fin, tail-plane
HLFC Active Wings, fin, tail-plane
Ribelts Passive Base fuselage
LEBUs Passive Base fuselage
Synthetic Jet actuators Active Wings, afterbody regions
Vortex generators Active/Passive Afterbody regions
Micro-vortex generators Passive Wings, rear fuselage
Gaster bump Passive Fin
Slot suction Active Wing/fuselage junction
Winglets Passive Tail plane

Table 5.3: Tilt-rotor regions suitable for application of the mentioned devices.

niques.
106 5. A Survey of the Techniques for Drag Reduction
Chapter 6

Introduction to Intake
Aerodynamics

The fuel consumption of turboshaft engines depends not only on the output
power but also on the aerodynamic performance of air inlets and to a lesser extent,
of exhaust nozzles. Engine performance loss as compared to the ground bench
reference may be quite substantial, in the order of 5% or more on some helicopters.
In this chapter, the intake aerodynamics issue will be addressed. The reason for
this is to conceptually contextualize the optimization problem presented in Part I
of chapter 7. Moreover, in section 6.4.1, the intake performance parameters and the
way they are evaluated for the AW101 Intake Test case will be addressed.

6.1 Preliminar considerations


From the point of view of the propulsive system adopted, an helicopter is
equivalent to a turboprop-shaft driven aircraft.
The main difference with the turboprop-driven aircraft lies in that the exhaust
velocity is negligible, whereas all the output power is required to drive the main
rotor shaft.
As others air-breathing jet engines, turboshafts must be supplied with air from
the atmosphere in which the aircraft is operating. The intake, or inlet, is that part of
the aircraft structure which performs such supply duty. Its aerodynamic behaviour
can be roughly understood by analyzing the progress of two stream flows from an
undisturbed condition far ahead of the aircraft.

• Internal flow: it enters the intake and the aerodynamicist’s task is to maxi-
mize the condition of pressure and flow uniformity with which it arrives at
the engine face. These properties are vital to the performance and stability of
engine operation.

• External flow: it passes around the intake as part of the general airflow over
the aircraft itself. Although not quantitatively definable in the manner of
the internal flow, the external flow is none-the-less important to the aircraft

107
108 6. Introduction to Intake Aerodynamics

aerodynamics, principally to aircraft drag, and its effect must be assessed by


suitable difference procedures.

The subject of intake aerodynamics is the study of both the internal and the
external flow as defined in this way, with the scope of quantify their influence on
the aircraft efficiency and performance. The system requirements, for an aircraft
intake design, depend on the aircraft mission specification.
Hereafter, the main design goals for intake aerodynamics are summarized:

• To provide the engine with adequate mass flow rate, at the proper Mach
number, at the engine face throughout the whole flight envelope;

• To guarantee a smooth and uniform feed flow into the engine compressor;

• To integrate well within the nacelle or the fuselage, leading in a low installa-
tion drag.

Nevertheless, when addressing intake design, also non-aerodynamic matters


have to be taken into consideration like, for example:

• To provide acoustic absorption of fan/engine noise;

• To provide a particles separator to avoid Foreign Object Damage (FOD);

• To be characterized by low weight and low cost of manufacture;

• To provide favorable accessibility, reliability and reparability.

The next chapter will deal with the topic of subsonic inlets, and both internal
and external flow patterns will be considered. The importance of intake efficiency
of both these two kinds of flow patterns will be pointed out.

6.2 Subsonic inlets


An engine installed in an aircraft must be provided with an air intake and a
ducting system. In Figure 6.1, a CAD drawing of the intake system, along with
particle separator and exhaust duct, is illustrated. The sketch refers in particular to
the intake installed into the wing-mounted engine of ERICA tiltrotor (see Figure
5.1).
At this point, it is also useful to illustrate the main features of a turbo-shaft
air feeding system. For turboshaft engines, the airflow entering the compressor
must have low Mach number, in the range 0.4 to 0.7, being the upper part of
the range suitable for transonic compressors only. If the engine is designed for
subsonic cruise, the inlet must be designed to act as a diffuser with reasonably
gentle diffusion from flight Mach number to a lower Mach number.
The inlet must be designed to prevent boundary layer separation, even when
the axis of the intake is not perfectly aligned with the streamline direction far
upstream of the inlet. In other words, the performance of the inlet must not be
8.1. Subsonic inlets

An engine installed in an aircraft must be provided with an air intake and a ducting system.
In Fig.2, a CAD drawing of the intake system, along with particle separator and exhaust duct, is
6.2 Subsonic inlets 109
illustrated. The sketch refers in particular to the new Tiltrotor concept ERICA (see Fig. 1).

Fig. 2 - CAD model of the ERICA intake and exhaust system


Figure 6.1: CAD model of the ERICA intake and exhaust system.

For turboprop engines the airflow entering the compressor must have low Mach number, in the
range 0.4excessively sensitive
to 0.7, being to pitch
the upper part (up-and-down)
of the range suitableand yaw (side-to-side)
for transonic motions only.
compressors of theIf the
aircraft. Moreover, this is one of the most challenging goal in intake
engine is designed for subsonic cruise, the inlet must be designed to act as a diffuser with aerodynamic
design,
reasonably gentlebecause
diffusionoffrom
theflight
adverse
Mach pressure
number gradient within
to a lower Machthe diffuser duct, which
number.
increases the attitude of the boundary layer to separate.
The inlet must be designed to prevent boundary layer separation, even when the axis of the
As a matter of fact, one recognizes that in a real flow environment, boundary
intake is not perfectly aligned with the streamline direction far upstream of the inlet. In other words,
layers are formed, and have the tendency to separate when exposed to a rising
the performance of the inlet must not be excessively sensitive to pitch (up-and-down) and yaw
static pressure, known as an adverse pressure gradient. Therefore, the behaviour
(side-to-side) motions of the aircraft.
of a diffuser is expected to be driven by the viscous region near its walls, i.e.,
As a matter of fact,
the state of theone recognizes
boundary thatasinina attached,
layer real flow environment,
separated, orboundary layers
transitory are formed,
(unsteady)
and haveconditions.
the tendency to separate when exposed to a rising static pressure, known as an adverse
pressure gradient.
The mainTherefore, the behaviour
geometrical features of of aa subsonic
diffuser isdiffuser
expected are:to be driven by the viscous
Questo documento
1. contiene dati ed informazioni
The shape of thedicross
proprietàsection;
Agusta che ne vieta la divulgazione e la riproduzione parziale o totale senza la sua autorizzazione
scritta.
This document contains data and information which are the property of Agusta. Agusta forbids its circulation and reproduction, whether in part or in its entirety, without a written authorization from the
company.
2. The shape of the centerline;
AG-TEC168 Rev. 07
3. The stream-wise area variation.

The three basic geometries of interest are presented in Figure 6.2:


Another feature of a real diffuser is the geometrical shape of its centerline.
Often the engine face is hidden from an observer looking through the inlet. The
feature of a hidden engine face offers the potential of masking the radar reflections
off the engine face, which is advantageous in a stealth aircraft. Turboshafts, such
as those mounted on the helicopter whose intake is addressed in Chapter 7, due to
the presence of the propeller shaft, require an S-shaped subsonic diffuser duct to
channel air to the engine face. From the fluid dynamics point of view, a curved
duct induces a secondary flow pattern, which essentially sets up “pockets” of
swirling flow at the duct exit. Often these pockets of swirling flow occur in pairs
and are counterrotating. The existence of twin swirling flows through curved pipes
2. the shape of the centerline;

3. the stream-wise area variation.

Some example of classical cross sectional shape are provided in figure 3.5;
110
even tough modern aerodynamic design6.mayIntroduction to Intake
produce more Aerodynamics
complex shape like
the ERICA one (figure 3.4).

CHAPTER 3. Introduction to the Aerodynamic Problem


(a) 2D rectangular (b) Conical (c) Annular
Figure 6.2: Schematic drawing of subsonic diffuser geometry with straight centerline
vortices
(takenusually
from
Figure Schematic
occur indrawing
3.5:[28].) pairs ofand
subsonic diffuser geometry withThis
are counter-rotating. straight centerline
effect is due to
[4].
the centrifugal force that pushes the higher energy streamlines toward the outer
radiiisof theknown:
well bend,centrifugal
while theforces
low push
energy streamlines are in turn forced
towardstothemove
Also the centerline geometry may the
havehigher energy
different stream
shapes but,lines
in turbo-props,
inwards
outer[23].
such radii
Forthe
details
thoseofmounted
bendonabout
while anlow
the the
wing
experimental
energy
tips
investigation
of a streamlines areto
tilt-rotor, due
ofpresence
in the
turn this phenomenon
forced toofmove
the
see reference
propeller shaft, an S-shaped subsonic diffuser duct is required to channel air shown.
inwards [24].
(see In
Figure figure
6.3). 3.6 a schematic drawing of an S-shaped duct is to
the engine face. The curvature of the centerline induces a secondary flow pattern
in the duct which is characterized by swirling vortices at the duct exit; these

31

Figure Drawing
3.6:6.3:
Figure Schematicofdrawing
an S-shaped duct
of an S-duct with
with a schematic
two pockets representation
of swirling flows (known of
as the
swirling flow [4].
the secondary flow patter) generated by the bends (taken from [28].)

TheThe swirling
swirling flow
flowinduced
induced by thecurved
by the curvedductduct leads
leads in a variation
in a local local variation
of the of
relative flow
the relative flowangle at the
angle at compressor inlet and,
the compressor in and,
inlet severein
situations, it may produce
severe situations, it may
rotating
produce stall instability
rotating of the compressor
stall instability rotor.
of the compressor rotor.

3.2.2
6.3 Inlet Operating
Inlet Operating Conditions
Conditions
An Depending
aircraft intake generally have to operate with a wide range of incident
on the flight speed and the mass flow demanded by the engine,
stream
the conditions
inlet may havein dependence on the
to operate with flightrange
a wide speedofand/or
incidentthe massconditions.
stream flow required
by the engine. Typical streamlines patterns in two different subsonic
Figure 6.4 shows the streamline patterns for two typical subsonic conditions conditions
and
and the
thecorresponding
corresponding thermodynamic behavior are shown in figure 3.7.
thermodynamic path of an “average” fluid particle. During level
During level
cruise, the cruise, the
streamline streamlines
pattern pattern
may include may includeofsome
some deceleration deceleration
the entering fluid of
external to fluid
the incoming the inlet plane (as
external to depicted
the inletinplane
Figure(3.7-a).
6.4-a).
On Onthethe contrary,during
contrary, duringlowlow speed,
speed, high
high trust,
trust,operations
operations likelike
take-off and and
take-off
climb, the same engine requires more mass flow and the streamlines
climb, the same engine requires more mass flow and the streamlines in the in the pre-pre-
entry zone may resemble as in figure 3.7-b, which shows external acceleration of
the pre-entry stream.
In both cases there is an isentropic external change of thermodynamic state.
For given air velocities of the undisturbed flow (station a, far ahead of the aircraft)
and at the compressor inlet (station 2 ), external acceleration raises the inlet
velocity and in turn decreases the inlet pressure, thereby increasing the intensity
of the adverse pressure gradient along the diffuser. In severe situations this effect
6.3 Inlet Operating Conditions 111
3.2 Introduction to Intake Aerodynamics

(a) High speed or low mass flow (b) Low speed or high mass flow

Figure
Figure 6.4: 3.7: Typical
Typical streamstream line patterns
line patterns in subsonic
in subsonic inlets from
inlets (taken [5]. [43].)

high engine
entry zoneload
mayoperations,
resemble aswhich also leads
in Figure 6.4-b,to external
which deceleration
shows in normal
external acceleration
operating conditions.
of the pre-entry stream. In both cases there is an isentropic external change of
thermodynamic state. For given air velocities of the undisturbed flow (station a, far
ahead ofInternal
3.2.3 the aircraft)Flow
and at the compressor inlet (station 2), external acceleration
raises the inlet velocity and in turn decreases the inlet pressure, thereby increasing
The aerodynamic behavior of the internal flow in an air intake can be quali-
the intensity of the adverse pressure gradient along the diffuser. In severe situations
tatively described by means of diffusers theory because, like in diffusers behavior,
this effect may lead in stall of the diffuser because of boundary layer separation
in any section of the intake duct momentum decreases and pressure increases,
induced by adverse pressure gradient. Conversely, external deceleration reduce
without any external work supply.
the pressure rise across the diffuser and hence the boundary layer situation is less
Even though, only a little part of diffusers theories and experimental works can
critical. Therefore, the inlet area is chosen so as to minimize external acceleration
directly be applied to subsonic intake because the most of these works are focused
during high engine load operations, which also leads to external deceleration in
on maximum pressure recovery condition, which is usually associated with highly
normal operating conditions.
non-uniform exit velocities and the presence of some unsteadiness in the exit
flow. In subsonic aircraft intakes there are always stringent requirements about
6.3.1
flow Internal
steadiness flow
and uniformity at the compressor face. therefore, historically
intakeQualitatively
design practice does not use results of diffusers research but are strongly
the flow in the inlet behaves as though it were in a “diffuser”,
based on computer codes calculation followed by wind tunnel tests to asses inlet
which is a common element in fluid machinery.
performance under a wide range of operating conditions.
A better term might be decelerator, since the device is not primarily concerned
Nowadays, intake design procedures may obtain grate assistance by computa-
with molecular or turbulent diffusion, but in this work the traditional term dif-
tional fluid dynamics (CFD) analysis involving the solutions of 3D Navier&Stokes
fuser will be retained and defined to mean any section of a duct in which fluid
momentum decreases and pressure rises, no work being done.
33
112 6. Introduction to Intake Aerodynamics

Considerable experimental and analytical work has been carried out on cylindri-
cal (especially conical and annular) diffusers, but little of this is directly applicable
to subsonic aircraft inlets. The reason is that most of the work on diffusers focuses
on the conditions that are related to maximum pressure recovery, which is usually
associated with a highly non-uniform exit velocity profile and perhaps even with
some flow unsteadiness.
In typical subsonic aircraft inlets, there is a stringent requirement that the
flow velocity entering the compressor be steady and uniform. Consequently, inlet
design CHAPTER
does not3. depend
Introduction sotomuch on theProblem
the Aerodynamic results of diffusers research as on potential
flow calculations, coupled with boundary layer calculations and followed by wind
tunnel equations
testing[25].to assess inlet performance under a wide range of test conditions.
Figure 3.8 shown the three zones in which boundary layer separation may
Nowadays, intake
occur in a typical design
air intake procedure
plane section [5]: is assisted by Computational Fluid Dy-
namics (CFD)
• Separation of the external flow in zone one will be discussed inequations.
analyses involving 3D Navier-Stokes the next
In actual engine inlets, separation can take place in any of the three zones
section;

shown in• Figure


Separation 6.5. Separation
of the internal of place
flow may take the inexternal
either zones flow
two and in zone 1 may result from
three,
depending on the geometry of the duct and operating condition.
local high velocities and subsequent deceleration over the outer surface. This
Zone three may be scene of large adverse pressure gradients, since the flow
circumstance, which
accelerates around theleads
nose of to
the high nacelle
central body, drag,
and then will be
decelerates discussed in section 6.6.
as the
Separation on the internal surfaces may take place in either zone 2 or zone 3,
curvature decreases.

Figure 6.5: Possible


Figure locations
3.8: Possible location ofof boundary
boundary layer [5].
layer separation separation (taken from [43].)

depending
3.2.4 on the geometry
External Flow of the duct and the operating conditions. Zone 3 may
be the scene of already
As it has quitepointed
largeout, adverse pressure
intake design gradients,
requires a compromise since the flow accelerates
between
around the nose of the centre body, and then decelerates as the curvature decreases.
external and internal deceleration to avoid boundary layer separation in both
regions.
At high angles
Figure 3.9 of attack,
shows a typical all
flow three zones could
pattern characterized by largebe subjected
external decel- to unusual pressure
eration. Flowing over the lip of the inlet, the external flow is accelerated to high
gradients.
velocity, which leads in a low pressure region, adversely affecting the boundary
layer flow in two different ways:

6.3.2 External flow


1. For purely subsonic flow, the low pressure region must be followed by a
region of rising pressure in which the boundary layer may separate. Hence
a limiting minimum pressure Pmin , or equivalently a limiting maximum ve-
As it has
locity already pointed
umax , can be defined as limitout, intake
value to design
avoid boundary requires a trade-off between
layer separation
downstream the lip. This is the typical case of a tilt-rotor nacelle;
external and internal deceleration to avoid boundary layer separation in both
regions.
34 Figure 6.6 shows a typical flow pattern characterized by large external

deceleration. Flowing over the lip of the inlet, the external flow is accelerated
to high velocity, which leads in a low pressure region, adversely affecting the
boundary layer flow in two different ways:

1. For entirely subsonic flow, the low-pressure region must be followed by


a region of rising pressure in which the boundary layer may separate (if
6.3 Inlet Operating Conditions 113

pressure gradients are high enough); hence one might expect a limiting low
pressure pmin or, equivalently, a maximum local velocity umax , beyond which
boundary layer separation can be expected downstream. This is the case of a
typical Tiltrotor nacelle.

2. For higher flight velocities (or higher local accelerations), partially supersonic
flow can occur. Local supersonic regions 3.2 usually end to
Introduction abruptly in a shock,
Intake Aerodynamics
and the shock-wall intersection may cause a boundary layer separation.
2. For higher local accelerations, partially supersonic flow may occur. Local
Whatever the cause, boundary layer separation is to be avoided, since it results
supersonic regions usually end abruptly in a shock, and the shock-wall in-
in poor pressure recovery in the flow over the after portions of the aircraft or
tersection can cause boundary layer separation. In this case, a limiting
engine housing. This,number
Mach of course,
can would result
be defined likeinthe
a net rearward
value not to force or dragto
be exceeded onavoid
the
body. A simpleshock
one-dimensional,
wave effects. incompressible, model, introduced by Küchemann

Figure 3.9: one-dimensional


Figure 6.6: One-dimensional model for
model for external external deceleration
deceleration study(taken study [4].[28].)
from

Whatever the cause, boundary layer separation has to be avoided, since it


and Weber ([53]), is useful to understand the relationship between the external
results in an high drag force on the body of the aircraft.
flow behavior to theone-dimensional,
A simple extent of external deceleration:
incompressible, in reference
model, introduced toby Figure
Küchemann6.6,
the external
and cross
Webersection
[26], isofuseful
the intake grows tothe
to understand a maximum
relationship area Amaxthe
between , and the
external
body remains cylindrical
flow behavior to thedownstream to this point.
extent of external The control
deceleration: volume
in reference toindicated
figure 3.9,
extends far
the from thecross
external intake
sectionon of thethesides and
intake upstream
grows ends, crosses
to a maximum area Amaxthe inlet
, and the
body remains cylindrical downstream to this point.
at its minimum area Ai , passes over the inlet surface, and extends downstream,The control volume indicated
extends far from the intake on the sides and upstream ends, crosses the inlet at
far enough to consider an exit velocity equal to the undisturbed flight velocity
its minimum area Ai , passes over the inlet surface, and extends downstream, far
u a (Figureenough
6.6). Thus, all the external flow enters and leaves the control volume
to consider an exit velocity equal to the undisturbed flight velocity ua
with an axial
(figurevelocity u a , while
3.9). Thus, all thethe internal
external flowflowentersenters with the
and leaves thecontrol
same velocity
volume withua
and leavesanwith
axial avelocity
velocity i , assuming
ua ,uwhile one-dimensional
the internal flow enters with flow
the over
same the inlet.uaThe
velocity and
equation leaves
whichwith
describes the uflow
a velocity characteristics
i , assuming in this simplified
one-dimensional intake
flow over the inlet.model is
the follow: The equation which describes the flow characteristics in this simplified intake
model is the follow:  2
ui
Amax 1 − ua (1 − c Pmax )
=A1max + (1 − uuai )2 (1 − cP max ) (6.1)
Ai = 1 + sc Pmax (3.1)
Ai scP max
where s is a value dependent on the shape of the nacelle and cP max is the
maximum value of the pressure coefficient over the nacelle. Assuming s = 0.5,

35
114 6. Introduction to Intake Aerodynamics

whereCHAPTER
s is a value3.dependent onto
Introduction thetheshape of the nacelle
Aerodynamic Problem and c Pmax is the maximum
value of the pressure coefficient over the nacelle. Assuming s = 0.5, for purposes of
illustration, it is shown in Figure 6.7, the dependence of the size of the external
the variation of the frontal area ratio (Amax /Ai ), necessary to avoid boundary
surfacelayer
necessary to prevent
separation, external
in dependance boundary
on the layer
deceleration ratioseparation,
(ui /ua ) can for any given
be shown
ui
value of . 3.10.
in ufigure
a

6.7: Minimum
Figure Figure frontalfrontal
3.10: Minimum area area
ratioratio
for for
various
variousvalues of ccPPmax
values of (s==0.5
max (s 0.5ininequa-
equation
tion 3.1),[5].
6.1)(taken from [43].)

The main point


Looking at here is that the
this picture, it islarger the external
clear that the largerdeceleration (i.e., the smaller
the external deceleration
(anduiso the smaller the ratio ui /ua ), the larger must be size of the nacelle to avoid
the value ua ), the larger must be the size of the nacelle if one is to prevent excessive
boundary layer separation; nevertheless, an increase in nacelle size also increase
drag. Even in the absence of separation, the larger the nacelle, the larger the
the total drag even without boundary layer separation, so a compromise have to
aerodynamic
be found. drag on it. However, if the external deceleration is modest, its effect
on minimum Evennacelle
though size
this is quiet pertains
analysis small. The to ause of partial
simplified internal
picture of thedeceleration
real flow is,
of course, doubly effective in reducing maximum diameter because it permits a
around inlets, it shows that performance of an air intake depends on the pressure
gradient
reduction on both
in both the internal
Ai and Amax and external surfaces; the external pressure rise
Ai . To summarize,
is fixed by external deceleration
even though this analysis
and the frontal area ratio (Amax /Ai ), while the
pertains
to a simplified picture of the real flow around inlets, nevertheless, it shows that
internal one depends on the velocity reduction between the entry to the inlet
the performance
diffuser and ofthean inlet depends
compressor on thesize
face. Nacelle pressure
required gradient oncan
for low drag both internal and
be strongly
external surfaces.
dependent onThe external
external pressure
deceleration, rise
with theisobjective
fixed bytothe external
avoid boundarycompression
layer
Awith
and the ratio ( Ai ) of maximum area to inlet area. The internal pressure rise
separation max the smaller nacelle size possible.

depends In onthethecase of turbo-propeller


reduction of velocity aircraft, also the
between propeller
entry swirlinlet
to the effectdiffuser
on the and
boundary layer development have to be taken into account, since it influences
entry tothethe compressor.
attitude Nacelle
of boundary layer size required
to separate for low drag
or reattach. can beofquiet
As a matter strongly
fact, the
dependent on the degree of external deceleration.
As far as nacelle design is concerned, drag reduction may be achieved through
36
a hybrid laminar flow control (please read section 5.4.3).
Moreover, in realistic analyses, one must consider compressibility effects.
It is to be pointed out that main rotor swirl effect on the boundary layer de-
velopment has to be taken into account, since it intimately affects its attitude to
6.4 Intake-Engine Integration: Flow Distortion Issue 115

separate or reattach. As a matter of fact, the main rotor slipstream may act so as
to energize the boundary layer and, hence to prevent its separations. Moreover,
3D effects (characterizing the real flows) and boundary layer effects of a typical
intake-fuselage configuration of an helicopter are not negligible. Once again, Com-
putational Fluid Dynamics (CFD) analyses involving 3D Navier-Stokes equations
can effectively assist engineers during more advanced design steps.

6.4 Intake-Engine Integration: Flow Distortion Issue


As mentioned in sections 6.3 and 6.2, the air diffusion process, which occurs
in the intake system, must be accomplished with the least loss in total pressure,
with the best flow distribution attainable at the compressor face and with the least
amount of drag.
In particular, the level of flow distortion, which an intake creates at the engine
aspiration plane, affects the performance and stability of the compressor. The term
distortion means simply a non-uniformity in the flow, which is characterized by a
destabilizing impact on the compressor performance, by means of reduction of the
stability margin (of compressor or fan), potentially to the level of compressor stall
or engine surge.
A lot of types of flow distortion can be defined, most of them are carefully
described in reference [28].
The main sources of flow distortions, that quantifies the uniformity of the flow
distribution at the intake/engine interface, can be classified into the following
categories [28]:

1. Total pressure distortion pt (r, θ );

2. Flow angle distortion α(r, θ );

3. Secondary flow-swirl at the engine face Cθ (r, θ );

4. Total temperature distortion Tt (r, θ );

5. Entropy distortion s(r, θ );

6. Combinations of some or all of the above distortions.

The common feature of all different types of distortion is found in their desta-
bilizing impact on the compressor performance. For helicopter inlets applications,
the dominant distortion effect is due to the total pressure distortion. Steady state
total pressure distortion is a measure of spatial non-uniformity of total pressure
time average values, over the engine face.
Total temperature and entropy distortions act in parallel with the total pressure
one, except when hot gases from external sources are ingested (i.e. exhaust gasses
from other aircrafts). Flow angle and swirl distortion impact the compressor
behavior in parallel with total pressure distortion, with similar effects; it has been
demonstrated that, especially at low incidence flight conditions, total pressure and
116 6. Introduction to Intake Aerodynamics

swirl coefficient have the same trend (an increase of swirl leads to an increase of
total pressure distortion) [79]. Moreover, total-pressure distortion may be steady or
time-variant (“dynamic”) and in the latter case, it may be of the spatially uniform
(“buzz”) type or of a spatially non-uniform (“turbulence”) type. Consistent with
the concept of total-total pressure ratio as a representative mean of time-averaged
values of total-pressure across the engine face position, is the concept of distortion
as a measure of the spatial non-uniformity of those time-averaged values. This
will be referred to as “steady-state” distortion, notwithstanding the fact that where
large transverse gradients of pressure are present, the flow time-wise can at best
be no more than quasi-steady. However, gas turbine engine manufacturers provide
limiting values of steady state total pressure distortion, which have to characterize
inlets in order to guarantee good engine performance.
Therefore, computation of steady state total pressure distortion is considered
sufficient for the intake performance optimization, which is the purpose of this
document.

6.4.1 Total pressure distortion issue and Definition of Intake Perfor-


mance Parameters
The most common measure of intake efficiency is the intake total pressure
ratio (also referred to as pressure recovery). The total pressure ratio represents the
efficiency of the intake compression process which transforms the free-stream
kinetic energy into static pressure. The total pressure ratio definition adopted is the
following [79]:

Pf
ηT = (6.2)
P∞
where

• Pf is total pressure measured at the AIP;

• P∞ is the free stream total pressure. In this case, it is coincident with the total
pressure imposed at the Inlet boundary.

However, for the purpose of the present work, the absolute value of total
pressure losses
∆PT = P∞ − Pf (6.3)
was preferred to the total pressure ratio, for practical reasons due to the
optimization problem formulation, and specifically to the use of penalty functions
for functional constraints handling, as will be better specified in chapter 7.32.
Actually, the absolute values of total pressure losses along the intake, calculated
both in forward flight and hovering conditions, will be used for the formulation of
the multi-point optimization problem, as will be illustrated in chapter 7.32.
The total pressure loss can be evaluated form CFD simulations by means of
standard Fluent® post processing functions; in particular Pf can be determined
using a mass-weighted-average surface integral (see [7]) of total pressure over the
6.4 Intake-Engine Integration: Flow Distortion Issue 117

AIP surface, while


Figure P ∞ is Flow
3.13: a boundary
distortionsvalue imposed
induced at the
by boundary layervirtual wind
separation [7]. tunnel inlet
(please see Figure 7.3). Since those values are known, the computation of ∆PT is
straightforward.
4. distortions induced by the propeller swirl.
The air Steady
diffusion process, occurring in the intake system, must be accomplished
state total pressure distortions are described in details in the following
with thesection.
minimum loss in total-pressure and with the best attainable flow dis-
tribution at the intake exit plane. In particular, since aircraft gas turbine engines
operate 3.3.1
downstream
TotalanPressure
air intake Distortion
system, the level of distortion at the compressor
face affectsSteady
the performance and the stability of the compressor itself. Therefore,
state total pressure distortion is a measure of spatial non-uniformity
anotherofimportant
total pressure time average values, aspect
intake performance over theisengine
the AIP
face.flow distortion.

(a) Radial tip distortion. (b) Radial hub distor-


tion.

(c) Circumferential hub (d) Full-span distortion


distortion. of angular extent θ.

Figure 3.14: Different types of total pressure distortion [4].


Figure 6.8: Different types of total pressure distortion (taken from [28].)

In describing the total pressure distortion and its impact on compressor


39
per-
formance, the spatial extent of the spoiled flow is divided according to its radial
and circumferential extent, as shown in Figure 6.8. It is necessary to derive a
quantitative measure of distortion, from which both the quality of intake flow and
the tolerance of an engine can be judged. Distortion coefficients may be defined in
various ways; Rolls Royce use the following one:

Pf − Pϑ
DC (ϑ ) = (6.4)
qf

where Pf is the weighted area average total pressure at engine face, q f is the
corresponding mean dynamic head and Pϑ is the weighted area average total
pressure in the “worst” sector of the face, of angle ϑ. The sector “ϑ ” must be of
significant extent and 60° is usually regarded as a satisfactory minimum. Thus a
commonly used coefficient is DC (60): others which are also used are DC (90) and
DC (120) [79].
Pf − Pθ (3.2)
DC(θ) =
qf
e face, qf is the corresponding
where Pf is the mean total pressure at the engin
pressure in the ”worst” sector,
mean dynamic pressure and Pθ is the mean total
r must be of significant extent and
of angular extent θ, in the face. The base secto
118 y 6.
valueIntroduction
. Thus a commonly to Intake
used paramAerodynamics
eter
60◦ is usually regarded as a satisfactor
0 and DC120.
is DC60; others which are also used are DC9

urs and θ sector for definition of dis-


Figure 3.15: Illustration of total pressure conto
Figure 6.9: Illustration ofn total-pressure
tortio coefficient [8]. contours and ϑ sector for definition of distortion
coefficient (taken from [79].)
about the limit value of one of
Usually engine manufacturers give indication
e to work in the proper manner.
these certification parameters, to ensure the engin
Maximum DC (60) is adopted by Rolls-Royce as certification parameter for
inlet total pressure Perf
sorthe ormance
3.3. 2 Effedistortion
ct of Distorti andon it will
on Com be used
presas distortion performance
parameter throughout the whole document.
An example of wind tunnel test results, show
ing the effect s of inlet total
1 2 in figure 3.16.
q f =press is thetion
Vf distor
2 ρ f ure mean dynamic
on comp pressure
ressor perfo rman evaluated on the engine face; since,
ce, is prese nted
al opera ting line, the undistorted stall limit and two
strictly speaking, its definition is dependent
In that figure the norm on the CFD solution, the following
shown in dashe d
speed (100% and 93% of the design speed) are
conventions haveshaft
corrected been chosen:
lines represent instead, the stall limits corre spond ing to the
lines. The four solid
face. Based on the impact
distortion patterns simulated at the compressor
• ρ f four
= the ; themarg
ρ∞ stall free-stream density
in deterioration, the is used
total press in
ure place
distor of types
tion the AIP density;
can be sorted
on severe:
uding with the most,
in following way, staring with the less, and concl
• Vf = ρ∞ṁA f . The AIP velocity is considered equal to the ratio between the AIP
mass 1flowdirate
l h band i product of the free-stream density and AIP area.
di the

These assumptions make easier the numerical computation of the DC (60).


Moreover, they guarantee that every time a CFD calculation is performed during
optimization run, it has the same dynamic pressure (the mass flow rate ingested
by the engine being the same) of another one. Even though this definition is
strictly valid for incompressible flows only, it may be extended to the case under
consideration for determination of the reference velocity. The procedure for the
DC (60) computation using Fluent CFD results is outlined in Appendix A: DC (60)
computation Procedure.

6.5 Effect of Distortion on Compressor Performance


As an example of wind-tunnel test results, Figure 6.10 is presented, showing
the effects of inlet total pressure distortion on compressor stability.
The normal operating line, the undistorted stall boundary, and two corrected
shaft speeds of 100% and 93% design are shown in dashed lines. Four solid lines
correspond to the stall boundaries of the four distortion patterns simulated at the
compressor face via screens. From having the least to the most impact on the stall
margin deterioration, the culprits are identified as

1. radial hub;

2. radial tip;
2. radial tip distortion;

3. circumferential hub distortion;


6.5 Effect of
4. Distortion on Compressor
full-span circumferential Performance
distortion. 119

Figure 3.16: Effect of total pressure distortion on compressor stability [4].


Figure 6.10: Effect of inlet total pressure distortion on compressor stability (taken from
[28].) It can also be noted that an engine, subjected to full-span circumferential
distortion and operating at 100% corrected speed, works with zero stall margin.
The previous results are confirmed in figure 3.17, that shows the measured
3. circumferential hub;
effects of inlet distortions on the performance of a nine-stage axial compressor.
The measurements show that the stall line moves considerably to the right, when
4. full-span circumferential
the engine distortion
is subjected to inlet circumferential distortion, lowering the compressor
stall margin.
FiguresWe
respectively. 3.18 andnote
also 3.18 show
that the
thecompressor
full-spandelivery pressure
distortion at the
at the surgecorrected
100% line
for different types of distortion.
speed operates
Two at
keythe stall may
features boundary, i.e.,byzero-stall
be deduced margin.
these data:
The previous results are confirmed in Figure 6.11, showing the measured effects
1. as the angular extent of the spoiled sector is increased, there is an angle
of inlet distortion on the performance of a 9-stage axial compressor.
above which the exit static pressure changes little (figure 3.18). This angular
The measurements showed
extent is often referredthat thecritical
as the stall line
sectorwith circumferential inlet distor-
angle;
tion moved considerably to the right, a degradation in compressor stall margin. The
2. fixing the total angular extent of the distortion, the effect of subdividing it
general trends of compressor performance with different inlet distortions are eluci-
into different numbers of equal sections is shown in figure 3.19. The great-
dated throughesta effect
seriesonofthe
experiments
loss of peak undertaken
pressure rise isbyobserved
Reid ([71]).
when A representative
there is only
set of these results is shown in Figure 6.12, which shows the compressor delivery
pressure at the surge line for different types of distortions. Two key aspects may
41
be deduced from these data:

1. As the angular extent of the spoiled sector (low inlet total pressure) is
increased, there in an angle above which the exit static pressure changes little
(Figure 6.12). This angular extent is often referred to as the critical sector
angle.

2. Fixing the total angular extent of the distortion, the effect of sub-dividing it
into different number of equal sections is shown in Figure 6.13. The greatest
effect on the loss of peak pressure rise is observed when there is only one
region. This suggests that inlet distortion patterns, which have a longer length
scale and a lower circumferential harmonic content, are the most important.
CHAPTER 3. Introduction to the Aerodynamic Problem
120 6. Introduction to Intake Aerodynamics
one region affected by distortions. This suggest that inlet distortion pat-
terns, which have a longer length scale and a lower circumferential harmonic
content, lead in an higher degradation of engine performance.

CHAPTER 3. Introduction to the Aerodynamic Problem

one region affected by distortions. This suggest that inlet distortion pat-
terns, which have a longer length scale and a lower circumferential harmonic
content, lead in an higher degradation of engine performance.

Figure 3.17: Effect of circumferential inlet distortion on multi-stage axial compressor


Figure 6.11: Effects performan
of circumferential
ce [9]. inlet distortion on multistage axial compressor
performance (taken from [77].)

Figure 3.17: Effect of circumferential inlet distortion on multi-stage axial compressor


performance [9].

Figure 3.18: Effect of circumferential distortion sector angle on surge pressure ratio
[9].

42 Figure 3.18: Effect of circumferential distortion sector angle on surge pressure ratio
Figure 6.12: Effect of circumferential
[9]. distortion sector angle on surge pressure ratio (taken
from [77].)

42
3.3 Intake-Engine Integ

6.5 Effect of Distortion on Compressor Performance 121


3.3 Intake-Engine Integration: Flow Distortion Issue

Figure 3.19: Effect of number of sector on surge pressure ratio [9].

Other researches identified a critical circumferential extent of the spoiled sec-


that causes the maximum loss inofthe stall pressure ratio of a compressor, is
Figuretor,
6.13: Effect Effect of sectors,
◦ of number of in
Figure 3.19: number sector
on on surge
surge pressure
pressure ratio [9].
ratio (taken from [77].)
at nearly 60 , as in evidence figure 3.20.
Other researches identified a critical circumferential extent of the spoiled sec-
tor, that causes the maximum loss in the stall pressure ratio of a compressor, is
at nearly 60◦ , as in evidence in figure 3.20.

circumferential spoiling on compressor performance[4].


Figure 6.14: Effect
Figure of Effect
3.20: of extent
the extent ofofcircumferential spoiling on compressor performance
Figure 3.20: Effect of extent of circumferential spoiling on compressor performance [4].
(taken from [28].)

Other research identified a critical circumferential extent of the spoiled43sector


that causes the maximum loss in the stall pressure ratio of a compressor43 is at nearly

60°, as evidenced in Figure 6.14.


122 6. Introduction to Intake Aerodynamics
Chapter 7

Results and Discussion

In this chapter three test cases will be considered in order to demonstrate the
strength and versatility of the optimization loop designed during my PhD course.
The first test case concerns the aerodynamic shape optimization of the AW101
left air intake. Two objective functions will be optimized by means of the Pareto
approach, that is the total pressure losses during forward flight and hovering
flight conditions. Functional constraints on the maximum total pressure distortion
allowed at the compressor inlet increase further the optimization complexity.
Commercial code Fluent was adopted as the CFD solver, and the optimization was
run in Windows environment.
The second test case concerns the aerodynamic shape optimization of the
ERICA nose region. One objective function was considered, that is the drag force
during forward flight condition. Ansys Fluent® was chosen as the CFD solver.
The third test case addresses the aerodynamic shape optimization of the ERICA
nose region, this time by exploiting OpenFOAM® as the CFD solver. Visibility
requirements were fulfilled, by properly setting up the design variables and
limiting the maximum allowable deformation displacement. This case differs from
the previous one with regard to both the mesh model and the CFD simulations
features.

123
124 7. Results and Discussion
Part I

Test case A:
Aerodynamic shape optimization
of the AW101 air intake ]1

125
127

7.1 Introduction

Specifically, the present work deals with the application of a state-of-the-art


optimization methodology to the shape optimization of the full scale AW101
engine air intake ]1 mounted on a flat plate, with the final aim of reducing the
total pressure losses in both forward flight and hover conditions.
The optimization problem considered was of the multi-point type, since the
total pressure losses occurring in both forward flight and hover conditions were
requested to be minimized simultaneously. In addition, the optimization problem
featured some constraints, both of the functional and geometrical type. Specifically,
the functional constraints were related to the necessity of maintaining the flow
distortions at the Aerodynamic Interface Plane (AIP) within acceptable values:
a Penalty Function approach was used to deal with this type of constraint. On
the other hand, the requirement that the final intake design is both able to be
manufactured and installed into the aircraft led to the definition of a series of
geometrical constraints that were considered in this preliminary optimization as
well.
First of all, an investigation was carried out in order to identify a proper CFD
model set-up for optimization purposes: specifically, the overall CPU time required
for optimization needed to be minimized while preserving the numerical accuracy.
For this reason, in a first step the effects of varying the total length of the dummy
duct located downstream the engine face on the flow field features at the AIP were
investigated. As a result, a proper length was identified which made it possible to
reproduce intake losses with an acceptable accuracy while featuring the smallest
number of 3D elements.
On a second stage, a sensitivity analysis of the CFD model to the computational
mesh refinement was performed. As a matter of fact, decreasing the number of grid
elements results in a reduction of simulation time, but, at the same time, the model
capacity to properly reproduce the intake flow characteristics is progressively
deteriorated. Hence, the goal of this second step was to identify a suitable model
resulting from a trade-off between these two aspects. In addition to that, the
selected CFD model had to be sufficiently robust for optimization purposes: it was
required to feature a stable convergence pattern, especially considering the large
deformations it undergoes during optimization.
Once the final CFD model was set up, the geometrical parameterization was
carried out and the geometrical constraints were implemented.
Finally, the air intake aerodynamic shape optimization was carried out and the
results of optimization in terms of achieved margins of improvement with respect
to the baseline were discussed in detail.
Results presented here are owned by the Cleansky Partner Consortium HEAVYcOPTeR
(led by University of Padova), aimed at studying enhanced engine installations for
improved performance and lower noise specifically designed for heavy helicopters
configuration.
128 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

7.2 Optimization Procedure


In this chapter, the whole sequence of operations necessary to carry out the
optimization study is described.
As already mentioned, the optimization procedure was not performed in a
single step; instead, multiple steps were required to obtain reliable results, while
reducing the overall optimization time as much as possible. Each operation will be
fully analyzed in the next chapters.
The whole sequence of operations carried out to address a multi-objective
optimization problem is shown in Figure 7.1. It is worth mentioning that each
block was carried out independently from the others. The upper portion of the flow
chart shows the set of preliminary operations that need to be accomplished before
starting the successive automatic loop. The procedure started from the elaboration
of a baseline geometry CAD model of the intake system of the AW101 intake ]1
provided by AgustaWestland Ltd in CATIA® format.
The following steps focused on the construction of the baseline CFD model: first
of all, the surface mesh was generated using Altair Hypermesh® ([3]) surface mesh-
ing tool, then the volume mesh was created by means of Ansys Tgrid®([8]), and
finally the CFD case was set up with the commercial CFD code Ansys Fluent®([7])
which is the standard commercial code for CFD simulation adopted at AgustaWest-
land.
Next, a study on aerodynamic pressure loss sensitivity to the grid features
was carried out as described in detail in chapter 7.8. In fact, it is well known that
dummy duct length and mesh refinement level greatly influence the computed
total pressure distribution over the AIP. Many combinations of these parameters
were tested to obtain a model characterized by an acceptable accuracy while
featuring the least possible number of cells. all, the surface mesh was generated
using Hypermesh® surface meshing tool ([3]), then the volume mesh was created
by means of TGrid®, and finally the CFD case was set up using the commercial
CFD code Ansys Fluent®([7]), which is the standard commercial code for CFD
simulation adopted at AgustaWestland.
Once this study was performed, the sensitivity analysis was concluded, and an
optimal mesh model identified. The selected model was “frozen” and submitted
to Fluent ® CFD code in order to define the “baseline” aerodynamic features. A
section will be dedicated to the baseline model. The latter was then imported
into Altair HyperMesh® environment for the geometrical parameterization and
constraining.
With the independent variables identified and the parametric model set up, the
optimization process could be finally addressed.
The post-processing of the optimization results refers to the lower part of the
flow chart in Figure 7.1. The output of the automatic optimization loop was the
optimal combination of the design variables of the individuals lying on the Pareto
frontier. The designer may identify the individuals of interest among the multiple
optimal solutions of the Pareto front and the meshed geometry of each selected
individual could be reconstructed using HyperMorph® ([3]).
7.3 The object of the optimization 129

The final step is the reverse engineering process, which allows obtaining the
deformed surface starting from the mesh; this process was performed using
CATIA®V5 following the steps reported in section 7.12.

Sensitivity
analysis

Geometry elaboration Surface mesh

Volume mesh

Parameterization

Optimal mesh
Parameterization
model

GeDEAII driven
optimization

Design variables Morphed surface


values mesh

CFD objective
Volume mesh
function values

Pareto front
optimal design
variables
combination

Optimal
Optimal
individuals mesh
geometries
reconstruction

Figure 7.1: Flow chart of the complete optimization procedure.

7.3 The object of the optimization


The scope of the present work was the optimization of the full scale AW101
engine intake]1, located at pilot left-hand side: the CATIA® model of the intake]1
mounted on the AW101 fuselage is illustrated in Figure 7.2. However, for the
purpose of the trial optimization which was carried out in the present task, the
130 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

Left air intake

Figure 7.2: CATIA® V5 models of AW101 model without main and tail rotors.

intake CAD model was greatly simplified: actually, the effects of the installation on
the AW101 fuselage were neglected in a first approximation, and the intake was
mounted over a flat plate, as depicted in Figure 7.3.

A close-up of the isolated engine intake]1 is illustrated in Figure 7.4, where


the AIP location is evidenced. As apparent, a dummy duct was included into
the intake model downstream the engine face, as it is usually recommended for
this kind of simulations. In fact, since the pressure distribution over the AIP is
one of the required outputs of the CFD computations, a pressure outlet boundary
condition can not be applied directly over the engine face without forcing the flow
field inside the intake to adapt to the imposed boundary condition in a somehow
artificial way. Hence, the output boundary condition needs to be applied over a
surface simulating the compressor face and located sufficiently downstream the
actual intake exit in order not to influence the flow quantities distribution over the
AIP. Obviously, the optimization carried out in the present work addressed only
the intake component, while the dummy duct was excluded from parameterization.

Two flight conditions were considered simultaneously for the intake optimiza-
tion, namely cruise forward flight and hover. The main features of the reference
flight conditions are summarized in Table 7.1.
7.3 The object of the optimization 131

�����������
����������
������

����������

������

����� ��������

Figure 7.3: Virtual wind-tunnel layout with the AW101 left air intake system installed on a
virtual flat plate.

Dummy duct Compressor


face

Air Intake

AIP

Figure 7.4: Close-up of the CAD model of the isolated air intake]1

Flight True air speed, True air speed, Pressure al- Static pressure, Compressor OAT,
Condition [Kts] [m/s] titude, [m] [m] mass flow rate [C ]
[Kg/s]
Forward Flight 120 61.73 609.6 (2000 94214 4.7219 284.19
[ f t ])
Hovering Flight 0 0 0 101325 5.6472 288.15

Table 7.1: Flight Conditions for Air Intake optimization.


132 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

7.4 Geometry elaboration


As already mentioned, the intake system model was made up of four compo-
nents, specifically the Air intake itself, the AIP, the dummy duct and the compressor
face, depicted in Figure 7.4. The intake system was installed onto a virtual flat
plate. The resulting model along with the virtual wind tunnel is depicted in Figure
7.3.
In Figure 7.5, a detailed view of the Air Intake and the Aerodynamic Interface
Plane is depicted: the AIP and compressor face are evidenced as well. The AIP
location is particularly important, since the intake performance, both in terms of
total pressure losses and flow distortions, were measured over this plane.

Air Intake

AIP

Figure 7.5: A detailed view of the Air Intake and the Aerodynamic Interface Plane.

In Hypermesh® environment, the model surfaces were organized into seven


distinct components, namely:

1. Inlet, which is depicted in Figure 7.3;

2. Outlet, which is depicted in Figure 7.3;

3. Symmetry, which is depicted in Figure 7.3;

4. Air Intake, which corresponds to the object of the optimization surface,


depicted in Figure 7.4;

5. Dummy duct, which corresponds to the cylindrical portion of the intake


system, depicted in Figure 7.4;
7.5 Mesh generation 133

Dummy duct

Air Intake

Compressor
face

Figure 7.6: A detailed view of the Air Intake with the dummy duct and the compressor
face.

6. AIP, which is depicted in Figure 7.5;

7. Compressor face, which is depicted in Figure 7.6;

8. Dummy flat plate, which is depicted in Figure 7.3.

7.5 Mesh generation


As far as the mesh generation is concerned, the software used to carry out the
meshing operations was respectively:
1. Hypermesh® surface meshing tool for the surface mesh generation;

2. Ansys Tgrid® for the volume mesh generation.


The choice of HyperMesh® is related to the fact that it was extensively used
during all the optimization activities. Therefore, it is worth to building up the mesh
models directly using this tool. HyperMesh® allows to import CAD geometries
in native as well as standard formats such as (IGES, STEP etc.), to clean up them
and to generate high quality surfaces meshes. The choice of Tgrid® is instead
related to flow solver selected to carry out the present work, that is Ansys Fluent®
. Tgrid® is in fact the standard volume meshing tool when Fluent® is adopted as
CFD solver.

7.5.1 Superficial mesh


The Hypermesh® surface meshing tool can be used to mesh the geometry
surfaces directly from the CATIA® format, after importing a step model previously
134 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

exported from CATIA® environment. Both the surface and the volume mesh sizes
are discussed in chapter 7.8, since they depend on both the level of surface grid
refinement and the virtual wind tunnel dimensions. The surface mesh was more
refined over the Air intake, which was the object of optimization, whereas the
superficial grid over the dummy duct and the flat plate was slightly coarser; finally,
the symmetry, inlet and outlet surface mesh was the coarsest one.

The choice of adopting different mesh sizes was motivated by the attempt to
contain the total amount of surface elements (and consequently the number of 3D
cells) as much as possible. The transition between two zones with different mesh
size was accomplished by properly selecting Hypermesh® Advanced meshing tool
parameters. This feature is a really important characteristic of a CFD mesh, and it
is fundamental for a good CFD solution convergence. Two different surface mesh
distributions on the Air intake surface were investigated during the sensitivity
analysis, as will be illustrated in chapter 7.8. The finally selected model was
characterized by the following features:

Mesh Type Triangular


Number of elements 15
Element size within the Air Intake [mm] 36738
Element size within the Dummy Duct [mm] 15 ÷ 50
Element size within the Dummy flat plate [mm] 10 ÷ 100
Element size inlet, outlet, and symmetry planes [mm] 100 ÷ 400
Element size within the AIP [mm] 11 ÷ 11
Maximum 2D skeweness angle [°] 52
Maximum 2D aspect ratio 2.65

Table 7.2: Finally selected number of elements and target element size for each fuselage
component.

In particular, the Surface deviation algorithm was used in Hypermesh® environ-


ment to create the superficial mesh over the Air Intake component. In fact, Surface
deviation is a meshing algorithm that allows HyperMesh to automatically vary
node densities and biasing along curved surface edges to gain a more accurate
representation of the surface being meshed. The maximum values of 2D skeweness
angle were reached on the dummy flat plate, whereas on the region of greatest
interest from a fluid dynamic point of view, the Air Intake component, the values
are by far smaller. A sketch of the finally selected surface mesh over the Intake
system component is reported in Figure 7.7, while the superficial mesh over the
virtual wind tunnel walls is illustrated in Figure 7.8.

In Figure 7.7, the element size transition between the Air Intake and the
compressor face is smooth, hence ensuring robustness during the following CFD
calculations.
7.5 Mesh generation 135

Figure 7.7: Superficial mesh over the Intake system.

Figure 7.8: Superficial mesh over the virtual wind tunnel walls.
136 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

7.5.2 Volume grid


Starting from the surface mesh described above, the volume mesh was gener-
ated by means of the CFD meshing tool Ansys Tgrid®. Actually, the surface mesh
created within Hypermesh® can be saved the Fluent format (.cas) which can be
easily read by Tgrid®. Tgrid® is in fact the standard volume mesher when Fluent®.
In Tgrid® environment, the volume between the intake system surface and the
wind-tunnel walls was filled with tetrahedral elements, while the boundary layer
region was modelled using prismatic cells. In Table 7.3, the Tgrid® settings selected
to generate the internal volume mesh are presented. Seven prismatic layers were
built over the intake system in order to ensure that the simulated boundary layer
was entirely contained within the prismatic layers over the whole fuselage, with
the exception of regions of separated flow. Moreover, a height of 1 [mm] was
given to the first prismatic layer in order to meet the y+ requirements for the wall
function calculation during the CFD simulation [7]. The Aspect Ratio17 B.L. Offset
method was used, since it allows to control the aspect ratio of the prism cells
that are extruded from the base boundary zone. This was a crucial point in the
Volume mesh generation, since there is(31/10/11)
HC/HIT09/WP1.2/D1/A a change of element size starting from
the intake system andofproceeding
Assessment towards
the optimization theformulation
problem wind tunneland walls. Therefore, this
method allowed to keepapplication
small theto a trial problem
gap between
(HEAVYcOPTer D1 part II)
the last prismatic cell size and that
of the first tetrahedral element, in all of the model regions. This is clearly visible
in Figure 7.10. Tgrid® journal file for volume mesh generation, containing all the

Figure 7.9: Longitudinal view of the whole volumetric mesh on a center line plane.
Figure 8: Longitudinal view of the whole volumetric mesh on a center line plane.
settings and the coordinates of refinement regions, can be found in Appendix C.4.
A longitudinal view of the whole volume mesh is depicted in Figure 7.9, while
Figure 7.10 illustrates a close-up of the mesh near the intake system surface. Finally,
a particular of the boundary layer near the compressor face is reported in Figure
7.11.
The Tri/Tet Growth rate parameter was set to 1.1, in order to obtain a fine and
uniform mesh in the regions of interest.

7.6 Fluid-dynamic model set up


A pressure-based solver
Figure 9: Close-up type
of the with
volume absolute
mesh near thevelocity formulation and steady
intake system.

approach was adopted for the tiltrotor simulations. The κ − ω SST model was
17 The aspect ratio is defined as the ratio of the prism base length to the prism layer height.
7.6 Fluid-dynamic model set up 137

Components with boundary layer Air Intake, Dummy Duct, Dummy Flat Plate
Components without boundary layer Inlet, Outlet, Symmetry, AIP, Compressor face
B.L. Offset method HC/HIT09/WP1.2/D1/A (31/10/11)
Aspect ratio (10)
B.L. Growth method Geometric
Assessment of the optimization problem formulation and
application to a trial problem
B.L. Number of layer (HEAVYcOPTer D1 part II) 5
B.L. First height 1
B.L. Growth rate 1.2
Tri/Tet Improve surface mesh option Enabled
Tri/Tet Refinment method Adv/front
Tri/Tet Cell size function Geometric
Tri/Tet Growth rate 1.1
Tri/Tet Refinement regions No

Table 7.3: Ansys Tgrid® settings for volume mesh generation.


Figure 8: Longitudinal view of the whole volumetric mesh on a center line plane.

Figure 9: Close-up of the volume mesh near the intake system.


Figure 7.10: Close-up of the volume mesh near the intake system.

selected for turbulence treatment.


The air was treated as an ideal gas having constant specific heats, which
automatically enables the equation energy resolution. This makes it possible to
include the compressibility effects in the numerical simulations: in fact, as it can be
deduced from Table 7.1, the forward-flight condition selected for the optimization
features a moderately high Mach number. In Table 7.4, the air properties assigned
for the intake nose optimization are summarized.
As far as the solution algorithm is concerned, a COUPLED scheme was adopted,
which solves the pressure and moment equations simultaneously. The Pseudo
Transient option
Figure 10:wasClose-upactivated,
of the volume which is a form
mesh: boundary layerof nearimplicit
the compressor under-relaxation
face. method
introduced
20 / 69 Thisin the islast
document Fluent®
the property release
of the author(s) (please
organization(s) and shall not read the
be distributed Theory
or reproduced withoutGuide
their [7]). This
formal approval.
option allowed to decrease enormously the number of required iterations, without
increasing the time and computational effort needed per each iteration. Hence, the
total computational time of each simulation was substantially reduced.
A Second Order Upwind discretization scheme was considered sufficient for
this work purposes. It was selected for all the variables, since it guarantees a high
accuracy of the numerical solution, due to its potential to improve sufficiently
spatial accuracy by reducing numerical diffusion, particularly for complex three-
800-91-067

138 Part I - Aerodynamic


Numerical shape optimization
optimization of the EH101 of the AW101
Rev. A air intake ]1
Pag. 23 of 86

Figure 11 - Close-up of the volume mesh: boundary layer near the compressor face.
Figure 7.11: Close-up of the volume mesh: boundary layer near the compressor face.

7.3 Fluid-dynamic model set up


A pressure-based solver type with absolute velocity formulation and steady approach was adopted
for the tiltrotor simulations. The k − ω SST model was selected for turbulence treatment.

dimensional flows, while not negatively affecting the total time requested for
The air was treated as an ideal gas having constant specific heats, which automatically enables
the equation energy resolution. This makes it possible to include the compressibility effects in the
simulations when
numerical compared
simulations: toit can
in fact, as thebethird
deducedorder upwind
from Table 1, the flightscheme.
condition selected for
optimization features a moderately high Mach number. In Table 4, the air properties assigned for
the intake nose optimization are summarized.

In particular, for the gradient spatial discretization, the Green-Gauss node


As far as the solution algorithm is concerned, a COUPLED scheme was adopted, which solves the
pressure and moment equations simultaneously. The Pseudo Transient option was activated,
based discretization
which is a form of scheme was chosen
implicit under-relaxation methodinstead
introducedof the
in the lastmore
Fluent® common
release (pleaseGreen-Gauss
read the Theory Guide [4]). This option allowed to decrease enormously the number of required
cell basediterations,
or Least withoutsquares
increasing thecell based
time and schemes,
computational because
effort needed per eachit is more
iteration. Hence, suitable for
the total computational time of each simulation was substantially reduced.
unstructured tetrahedral
A Second mesh [4],scheme
Order Upwind discretization as iswasthe case sufficient
considered for thefor intake system
this work purposes. It simulation.
The selected solver settings and discretization schemes are summarized
was selected for all the variables, since it guarantees a high accuracy of the numerical solution,
due to its potential to improve sufficiently spatial accuracy by reducing numerical diffusion,
in Table
7.5. particularly for complex three-dimensional flows, while not negatively
requested for simulations when compared to the third order upwind scheme.
affecting the total time

Regarding the under-relaxation factors, they were left to their default values.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
As regardsInternational
the baseline run,
Ltd. (Collectively the
known solution was initialized by means of the Hybrid
as “AgustaWestland”)

Initialization approach [7]. Hybrid initialization is a collection of recipes and


Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

boundary interpolation methods. It solves Laplace’s equation to determine the


velocity and pressure fields. All other variables, such as temperature, turbulence,
species fractions, volume fractions, etc., will be automatically patched based on
domain averaged values or a particular interpolation recipe.

As far as the morphed cases is concerned, an interpolation file was used to map
data from the baseline mesh onto the deformed one. To do so, an interpolation file
was created for both the flown conditions. These files contained all the needed fluid
dynamic variables regarding the baseline CFD run. This strategy was followed in
order to decrease further the computational time, since, reasonably, the morphed
cases feature a fluid dynamic variables distribution not so different from the
baseline one.

For each simulation, the convergence criterion was established when normal-
ized RMS residuals were less than 1 · 10−6 . Furthermore, the Mass weighted
average of the total pressure on the AIP was monitored, in order to make sure it
would reach stabilized values at the end of the simulation.
7.7 Boundary and operating conditions 139

Gas thermo-dynamic model Ideal Gas


Cp Specific Heat [ J/kg · K ] Constant: 1006.43
Thermal conductivity [W/m · K ] Constant: 0.0242
Viscosity variation law Sutherland: three coefficients methods
µ0 Reference viscosity [kg/m · s] 1.716 · e−5
T0 Reference temperature [K ] 273.11
S Effective temperature [K ] 110.56

Table 7.4: Air properties for the intake optimization.

Solver type Pressure-based steady-state


Pressure-velocity coupling scheme COUPLED (Pseudo-transient option activated)
Spatial discretization schemes
Gradient Green-Gauss node based
Pressure Second order
Density Second-order Upwind
Momentum Second-order Upwind
Turbulent kinetic energy Second-order Upwind
Specific dissipation rate Second-order Upwind
Energy Second-order Upwind

Table 7.5: Solver settings and discretization schemes.

7.7 Boundary and operating conditions


The boundary conditions for the optimization study were selected according to
the operating flight conditions reported in Table 7.1. A total pressure condition was
imposed on the wind tunnel inlet, while a static pressure was assigned over the
outlet section. Over the lateral walls of the wind tunnel, together with the bottom
surface, a symmetry condition was imposed. All the other surfaces were treated as
hydraulically smooth, adiabatic walls. On the compressor face patch, a pressure-
outlet boundary condition was imposed, by specifying a target mass flow rate,
according to the flown condition being simulated. The mass flow rate values are
specified in Table 1. The operating reference condition was set to 0 [Pa] for both the
investigated operating conditions. The pressure inlet boundary condition imposed
on the inlet boundary zone requires the specification of both the gauge total
pressure and the total temperature. These quantities were determined applying
the ideal gas laws to the data of Table 7.1; in fact, from the static temperature (T∞
in [K], corresponding to the OAT given in Table 7.1) and the velocity (V∞ , the True
air speed) the operating Mach number (Ma∞ ) was calculated as:

V∞
Ma∞ = √ (7.1)
kRT∞
where k is the specific heats ratio (1.4 for dry air) and R is the gas constant
(287 [ J/(kg · K )] for dry air). The resulting operating Mach number is 0.1826. The
following equations were used to calculate the total pressure and total temperature
140 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

at the inlet:
 k −1
k−1

k
PT = P∞ 1 + Ma2∞ (7.2)
2

k−1
 
TT = T∞ 1 + Ma2∞ (7.3)
2
being P∞ the static pressure given in Table 7.1.
Subtracting the value of the reference pressure to the total pressure calculated
with Equation 7.2, we obtained a gauge total pressure of 96433.255 [Pa] for the
forward flight, and 101325 for the hovering flight, while the resulting total tem-
perature from Equation 7.3 was 286.086 [K ] and 288.15 [K ], respectively. Table 7.6
summarizes the settings for the pressure inlet boundary condition.

Gauge total pressure [ Pa] 96433.255 (Forward Flight) 101325 (Hovering Flight)
Supersonic/initial gauge pressure [ Pa] 94214 (Forward Flight) 101325 (Hovering Flight)
Direction specification method normal to boundary
Turbulence specification method Intensity and length scale
Turbulent intensity [%] 1
Turbulent length scale [m] 0.5
Total temperature [K ] 286.086 (Forward Flight) 288.15 (Hovering Flight)

Table 7.6: Pressure inlet specification at inlet.

The pressure outlet boundary condition requires the specification of the gauge
static pressure and the back-flow total temperature at the outlet surfaces: a relative
zero gauge pressure (corresponding to the undisturbed value) and a back-flow
total temperature equal to the inlet static temperature were chosen. Table 7.7
summarizes the settings for the pressure outlet boundary condition on the outlet
surfaces.
Gauge pressure [ Pa] 94214 (Forward Flight) 101325 (Hovering Flight)
Direction specification method normal to boundary
Turbulence specification method Intensity and length scale
Turbulent intensity [%] 1
Turbulent length scale [m] 0.5
Backflow Total temperature [K ] 284.19 (Forward Flight) 288.15 (Hovering Flight)

Table 7.7: Pressure-outlet specification at outlet.

Regarding the turbulence specification method, a turbulence intensity of 1as


the turbulent length scale is concerned, the final selected value 0.5 [m] was the
result of an√iterative calculation process. In fact, the turbulent length scale can be
defined as κ/ω.
Hence, the values of κ and ω were iteratively computed at both inlet and outlet
on a CFD model characterized by an initial turbulent length scale of 1 [m]. However,
some tests were performed that allowed to verify that the tiltrotor nose drag was
7.8 Grid sensitivity analysis 141

quite insensitive to variations of the turbulent length scale, at least for reasonable
values of this parameter. As far as the compressor face boundary condition is
concerned, a preliminary consideration has to be done. Intake flow field is strongly
dependent on the mass flow rate imposed by the engine compressor at the AIP.
The value of AIP mass flow rate is usually a function of both flight operating con-
ditions (flight speed, altitude, external pressure and temperature etc.) and power
demanded by the rotor. The compressor aspiration effects can be well modeled in
Fluent® by imposing at the AIP a pressure-outlet boundary condition with target
mass flow rate specification. Such a kind of boundary condition iteratively adjusts
the static pressure on the outlet surface until the mass flow rate measured on the
surface itself matches the specified value [7]. Unfortunately, the problem with this
solution is the constant static pressure distribution which is imposed on the outlet
surface by the boundary condition. This is not acceptable at the AIP, which is the
location where inlet performance has to be measured. For this reason, an additional
portion of the engine duct, behind the AIP, was modeled with the purpose of
assigning the target mass flow boundary condition on a surface downstream the
AIP. This additional dummy duct is clearly visible in Figure 7.6. Of course, the
flow field characteristic over this additional portion of the engine duct will be not
considered since the geometry results from a series of assumptions. However, it
allows to unconstraint the AIP static pressure distribution, resulting in a more
suitable CFD model for the intake system flow field. The location of the compressor
face was established by means of a sensitivity analysis presented in section 7.8. On
the compressor face, the pressure-outlet boundary condition features the following
settings.

Gauge pressure [ Pa] 94214 (Forward Flight) 101325 (Hovering Flight)


Direction specification method normal to boundary
Turbulence specification method Intensity and length scale
Turbulent intensity [%] 1
Hydraulic diameter [m] 0.5
Backflow Total temperature [K ] 284.19 (Forward Flight) 288.15 (Hovering Flight)
Target mass flow rate [kg/s] 4.7219 (Forward Flight) 5.6472 (Hovering Flight)

Table 7.8: Pressure-outlet specification at compressor face.

7.8 Grid sensitivity analysis


In this section, a preliminary sensitivity analysis is described, which was carried
out in order to identify a suitable CFD model to be used during the intake trial
optimization. Actually, the final selected CFD model must be a trade-off between
numerical accuracy and required computational time and resources. As stated
before, the final selected computational mesh was rather coarse: in fact, the scope of
the present work was to become familiar with the problem of the AW101 air intake
optimization and to demonstrate the capability of the proposed optimization
142 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

methodology to successfully handle such a problem with its specific objective


functions and constraints.
To this purpose, the following strategy was followed:

1. First, the dummy duct length was varied in order to point out the influence
of the duct length on the pressure losses and the DC (60) parameter in
both the flight conditions considered.The sensitivity analysis to the dummy
duct length was carried out using a coarse computational mesh, whose
characteristics are discussed later on.

2. Once the proper duct length area was selected, the influence of the grid
refinement was highlighted.

As far as the dummy duct length is concerned, three different values were
analyzed (namely 500 [mm], 1,200 [mm] and 2,000 [mm]) while keeping all the
other dimensions of the fluid domain fixed to their original values, and the effects
on the intake performance were registered, both in terms of total pressure losses
and maximum DC (60), for both forward flight and hover conditions.
The results of the sensitivity study are summarized in Table 7.9.

Dummy duct Total pressure Total pressure DC (60) max- DC (60) max-
length [mm] losses. losses. imum value. imum value.
Forward Flight Hovering Flight Forward flight Hovering flight
[ Pa] [ Pa]
500 831 200 0.461 0.027
1200 801 218 0.427 0.041
2000 802 226 0.403 0.044

Table 7.9: Total pressure on AIP sensitivity to dummy duct length.

As apparent, the total pressure losses in forward flight condition seem to reach
an asymptotic value with increasing length of the virtual duct, and in particular
they appear already stabilized for a dummy duct length equal to 1,200 [mm].
On the other hand, regarding the hovering conditions, the total pressure losses
tend to increase with increasing length of the dummy duct, even though the
maximum variation is within 10% of the baseline. Moreover, the DC (60) maximum
value tends to decrease as the duct length increases in forward flight conditions,
while the opposite occurs in hover conditions. However, excursions of DC (60) are
considered within acceptable values, at least for the duct lengths of 1200 [mm]
and 2000 [mm], for both forward flight and hover conditions. For the sake of
completeness, in Figure 7.12 and Figure 7.13 a graphical representation of the AIP
DC (60) sensitivity to dummy duct length is illustrated for the forward flight and
hover conditions respectively.
As apparent, the predicted DC (60) distributions at the AIP using a dummy
duct length equal to 1200 [mm] and 2000 [mm] respectively (corresponding to
red and black dotted lines in the figures), are very close to each other in both the
considered flight conditions. For this reason, the second model was adopted since
7.8 Grid sensitivity analysis 143

0.5
dummy duct length 500 [mm]
0.4
dummy duct length 1200 [mm]
0.3
dummy duct length 2000 [mm]

0.2
DC60(θ)

0.1

-0.1

-0.2

-0.3
0 50 100 150 200 250 300 350
θ [°]

Figure 7.12: AIP DC (60) sensitivity to dummy duct length. Forward Flight condition

dummy duct length 500 [mm]

0.1 dummy duct length 1200 [mm]

dummy duct length 2000 [mm]

0.05
DC60(θ)

-0.05

-0.1

0 50 100 150 200 250 300 350


θ [°]

Figure 7.13: AIP DC (60) sensitivity to dummy duct length. Hovering Flight condition
144 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

it appeared a good compromise solution for limiting the total amount of mesh
elements.
In a second step, a grid refinement was carried out on the selected geometrical
configuration, with the final aim of increasing the robustness of the CFD model
for optimization purposes.
Actually, the objective here was not to implement a highly accurate numerical
model fully validated against experimental data, but rather to set up a sufficiently
reliable and robust computational grid, suitable for use in the optimization runs.
For this reason, the final selected mesh was rather coarse.
A comparison of the characteristic features of the two analyzed grids is reported
in in Table 7.10.
Coarse Model Refined model
Mesh type Triangular Triangular
Number of elements 29941 37897
Element size within the Air Intake [mm] 16 ÷ 50 15
Element size within the Dummy duct [mm] 20 ÷ 200 15 ÷ 50
Element size within the Dummy Flate plate [mm] 50 ÷ 400 10 ÷ 100
Element size within the inlet, outlet and symmetry planes [mm] 150 ÷ 400 100 ÷ 400
Element size within the AIP [mm] 16 ÷ 50 11 ÷ 11
Maximum 2D skeweness angle [°] 67 52
Maximum 2D aspect ratio 4.59 2.65

Table 7.10: Superficial mesh features of the coarse and refined model.

Moreover, the effects of the grid refinement on the air intake performance, both
in terms of total pressure losses and maximum DC (60), are summarized in Table
7.11.
Mesh Total pressure Total pressure DC (60) max- DC (60) max-
refinement losses. losses. imum value. imum value.
level Forward Flight Hovering Flight Forward flight Hovering flight
[ Pa] [ Pa]
Coarse 801 218 0.43 0.04
Fine 794 207 0.45 0.002

Table 7.11: Total pressure on AIP sensitivity to mesh refinement.

Apparently, as far as the total pressure losses are concerned, the percentage
variation was negligible in forward flight conditions (<1%), whereas it became
larger in hovering flight (≈ 5%). Also regarding the DC (60) maximum values,
a remarkable percentage variation was evidenced in hover conditions, while the
forward flight value was less sensitive to the grid refinement level.
On the basis of the sensitivity analysis results, the refined mesh was selected for
the subsequent optimization runs: actually, no further refinement was investigated,
since the model was considered accurate enough for the trial optimization purposes.
Moreover, it was also more robust than the coarser one from a computational point
HC/HIT09/WP1.2/D1/A (31/10/11)
Assessment of the optimization problem formulation and
7.8 Grid sensitivity analysis
application to a trial problem 145
(HEAVYcOPTer D1 part II)

of view, which is a crucial characteristic during the optimization phase.


7. AERODYNAMIC CHARACTERIZATION OF THE INTAKE BASELINE
GEOMETRY
7.8.1 Aerodynamic characterization of the intake baseline geometry
Once the final CFD model to be used for optimization was selected, the aerodynamic
Once the final CFD model to be used for optimization was selected, the aero-
features offeatures
dynamic the baseline AW101
of the intake#1
baseline AW101wereintake1
analyzed. In fact,
were the weaknesses
analyzed. andweak-
In fact, the the
major sources
nesses and theofmajor lossessources
occurring in the occurring
of losses baseline geometry could begeometry
in the baseline identified: could
this isbe
identified:
fundamental this is fundamental
to find to find out
out a proper geometry a proper geometry
parameterization parameterization
for a significant performancefor
aimprovement.
significant performance improvement.
Firstof of
First all,all,
thethe y+ characteristics
y+ characteristics of the of
CFDthemodel
associated CFD solution
were verified. In Figure were
15, theinves-
y+
tigated. In Figure 7.14, the y+ distribution over the intake walls
distribution over the intake walls is illustrated, for the forward and hovering flight conditionsis illustrated,
for the forward and hovering flight conditions respectively. The selected param-
respectively. As apparent, the selected parameters for grid generation were shown to
eters for grid generation were not shown to guarantee that the nondimensional
guarantee that the non-dimensional mesh thickness at the intake surface fell within the
mesh thickness at the intake surface strictly fell within the discretization levels
+
(y = 30 ÷ 300)
discretization ( y = 30 ÷ 300
levelssuggested for )the standard
suggested wall
for the functions
standard implemented
wall functions in the
implemented
conventional turbulence models to work properly, in particular
in the conventional turbulence models to work properly, in particular over the intake region, over the intake
region,
where the where the fluid-dynamic
fluid-dynamic quantitiesfor
quantities necessary necessary for the
the objective optimization
functions objective
evaluation are
functions evaluation are calculated.
derived. Nevertheless, the Automatic near-wall treatment is used as the default in Fluent®
Nevertheless, in Fluent 13 [7] the Automatic near-wall treatment is used as the
13.0 for all the turbulence models based on the ω-equation. This option automatically
default in all models based on the ω-equation. This option automatically switches
switches from wall-functions to a low-Reynolds near wall formulation, as the mesh is refined.
from wall-functions to a low-Reynolds near wall formulation, as the mesh is refined.
Hence,the
Hence, theCFDmodelsmodelareis relatively
relativelyinsensitive to the(as
insensitive local value of
clearly y + ([4]).
stated in [7]) to the local Y
plus.

Forward flight Hover

Figure 15: Contours of wall y+ over the air intake walls: forward flight (on the left) and hovering
Figure 7.14: Contours of wall y+ over (on
thethe
Airright).
Intake walls: forward flight (on the left) and
hovering (on the right).
34 / 69 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
The total
formal pressure contours over the air intake walls are illustrated in Figure
approval.

7.15 for both forward flight and hovering conditions.


The figure clearly highlights the presence of a concentrated total pressure loss
located in the elbow of the air intake (in both the flight conditions). Actually, in this
region the flow is accelerated, due to the highly curved shape of the duct; hence, a
HC/HIT09/WP1.2/D1/A (31/10/11)
Assessment of the optimization problem formulation and
146 Part I - Aerodynamic shapetooptimization
application a trial problem of the AW101 air intake ]1
(HEAVYcOPTer D1 part II)

Total pressure reduction


due to friction effects
Total pressure loss due
to flow separation

Forward flight Hover

Figure 16: Contours of total pressure ([Pa]) over the air intake walls: forward flight (on the left)
and hovering
Figure 7.15: Contours of total pressure (on thethe
([Pa]) over right).
air intake walls: forward flight (on
the left) and hovering (on the right).

Forward flight Hover


localized loss occurs which is mainly related to concentrated friction effects. In fact,
Strong
this flowsecondary and
feature needs to be taken into account when identifying the geometry
swirling flows
parameterization strategy: to this purpose, proper shape functions were selected
aimed at reducing this localized total pressure loss, as will be illustrated in section
7.9.
Massive flow
Moreover, a total pressure drop is evidenced over the external surface of the
separation
S-shaped duct in forward flight conditions, due to a massive flow separation
occurring in the entry region of the air intake, as can be visualized in Figure
7.16, where the streamlines pattern inside the air intake in both forward flight
and hovering conditions are illustrated. In fact, the above mentioned separation
results in a degradation of the intake performance in forward flight conditions.
Moreover, a vortex flows appears inside the intake close to the AIP, due to the
Figure 17: Streamlines pattern inside the air intake: forward flight (on the left) and hovering (on
strong deflection the flow experiences around the right).the engine shaft. Actually, this vortex
flow is propagated towards the AIP, and it is expected to negatively affect the total
The total pressure contours over the air intake walls are illustrated in Figure 16 for both
pressure distribution at the engine face, as will be observed later on. As apparent
forward flight and hovering conditions. The figure clearly highlights the presence of a
from Figure 7.16, the forward flight condition is by far the most critical from the
35 / 69 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
intake efficiencyapproval.
formal point of view: in fact, in hover conditions the flow field appears
to be very regular and no separation occurs inside the intake duct.
The above mentioned flow characteristics have a direct impact on the total
pressure distribution over the AIP, hence affecting the compressor inlet flow
distortions. The total pressure contours over the AIP in both forward flight and
hovering conditions are reported in Figure 7.17.
As apparent, in forward flight condition the vortex flow originating in the inner
portion of the intake duct results in a non-symmetric distorted flow over the AIP.
On the other hand, in hover conditions the wake originated from the sharp corner
on the upper part of the air intake (visible in Figure 7.4) promotes a symmetrical
7.8 Grid sensitivity analysis 147
Figure 16: Contours of totalHC/HIT09/WP1.2/D1/A
pressure ([Pa]) over the (31/10/11)
air intake walls: forward flight (on the left)
Assessment of the and hovering (onproblem
optimization the right).formulation and
application to a trial problem
(HEAVYcOPTer D1 part II)
Forward flight Hover
concentrated total pressure loss located in the elbow of the air intake (in both the flight
conditions).
Strong secondary Actually,
and in this region the flow is accelerated, due to the highly curved shape of
swirling
the duct;flows
hence, a localized loss occurs which is mainly related to concentrated friction
effects. In fact, this flow feature needs to be taken into account when identifying the
geometry parameterization strategy: to this purpose, proper shape functions were selected
Massive flowat reducing this localized total pressure loss, as will be illustrated in §8.2.
aimed
separation
Moreover, a total pressure drop is evidenced over the external surface of the S-shaped
duct in forward flight conditions, due to a massive flow separation occurring in the entry
region of the air intake, as can be visualized in Figure 17, where the streamlines pattern
inside the air intake in both forward flight and hovering conditions are illustrated. In fact, the
above mentioned separation results in a degradation of the intake performance in forward
flight conditions. Moreover, a vortex flows appears inside the intake close to the AIP, due to
Figure
the 17: Streamlines
strong pattern
deflection the inside the airaround
flow experiences intake: forward flight
the engine (on the
shaft. left) and
Actually, thishovering (on
vortex flow
Figure 7.16: Streamlines pattern inside the the airright).
intake: forward flight (on the left) and
is propagated towards the AIP, and it is expected to negatively affect the total pressure
hovering (on theTheright).
total pressure contours over the air intake walls are illustrated in Figure 16 for both
distribution at the engine face, as will be observed later on.
forward flight and
As apparent hovering
from Figure conditions. The flight
17, the forward figurecondition
clearly is
highlights themost
by far the presence of a
critical from
35 / 69 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
flow distortion in
the intake the upper
efficiency
formal approval. pointportion
of view: in of
fact,the AIP.conditions the flow field appears to be very
in hover
regular and no separation occurs inside the intake duct.

Highly non-symmetrical
flow due to the vortex Nearly symmetrical
Forward flight Hover distortions due to
generated inside the
intake the sharp corner

Figure 18: Total pressure distribution ([Pa]) over the AIP surface: forward flight (on the left) and
Figure 7.17: Total pressure distribution hovering
([Pa]) over the
(on the AIP surface: forward flight (on the
right).
left) and hovering (on the right).

36 / 69 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
The circumferential
formal approval.
distribution of DC (60) over the AIP for the intake ]1
baseline geometry is illustrated in Figure 7.18 and Figure 7.19 for forward flight
and hover conditions respectively. Finally, the location of the DC (60) worst sector
for the two considered flight conditions is illustrated in Figure 7.20.
For the forward flight condition, the maximum DC (60) has a value of 0.45, and
the worst sector is located at 340 [deg]. On the other hand, in hover condition the
maximum DC (60) is 0.0205, and the worst sector is located at 360 [deg]. Moreover,
a somehow more regular and “periodical” fluctuation of distortions along the
circumferential direction is evidenced in hover (as can be appreciated in Figure
7.19), than that occurring in forward flight condition. The initial assessment of
HC/HIT09/WP1.2/D1/A (31/10/11)
Assessment of the optimization problem formulation and
application to a trial problem
(HEAVYcOPTer D1 part
HC/HIT09/WP1.2/D1/A II)
(31/10/11)
Assessment of the optimization problem formulation and
application to a trial problem
The above mentioned flow characteristics have a direct impact on the total pressure
(HEAVYcOPTer D1 part II)
148 Part I -over
distribution Aerodynamic shape
the AIP, hence affecting the optimization
compressor inlet flowof the AW101
distortions. The totalair intake ]1
pressure
Thecontours over the AIP
above mentioned in characteristics
flow both forward flight
haveanda hovering conditions
direct impact on thearetotal
reported in
pressure
Figure 18.
distribution over the AIP, hence affecting the compressor inlet flow distortions. The total
DC60MAX
pressure contours over the AIP in both forward flight and hovering conditions are reported in
Figure 18.
DC60MAX
Forward flight

Forward flight

Figure 19: Circumferential distribution of DC60 for the intake#1 baseline model: forward flight.

Figure 7.18: Circumferential distribution of DC (60) for the intake ]1 baseline model:
forward flight.
Figure 19: Circumferential distribution of DC60 for the intake#1 baseline model: forward flight.

Hover

DC60MAX

Hover

DC60MAX

HC/HIT09/WP1.2/D1/A (31/10/11)
Assessment of the optimization problem formulation and
application to a trial problem
(HEAVYcOPTer D1 part II)

hover conditions theCircumferential


Figure 20: wake originated from the
distribution sharp
of DC60 forcorner on the
the intake#1 uppermodel:
baseline part of the air intake
hover
conditions.
(visible in Figure 4) promotes a symmetrical flow distortion in the upper portion of the AIP.
As apparent, in forward flight condition the vortex flow originating in the inner portion of
The circumferential distributiondistribution
Figure 20: Circumferential of DC60 of over the
DC60 forAIP for the intake#1
the intake#1 baseline
baseline model: hovergeometry
the intake duct results in a non-symmetric conditions.
distorted flow over the AIP. On the other hand, in
Figure 7.19: Circumferential
is illustrated distribution
in Figure 19 and Figure of DC
20 for forward (60and
flight ) for
hoverthe intake respectively.
conditions ]1 baseline model:
hovering flight.As apparent, in forward flight condition the vortex flow originating in the inner portion of
Finally, the location of the DC60 worst sector for the two considered flight conditions is
the intake duct results in a non-symmetric distorted flow over the AIP. On the other hand, in
37 / 69 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
illustrated in Figure 21.
formal approval.

37 / 69 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
Forward flight
formal approval.
Hover

Figure 21: Location of the DC60 worst sector for the intake#1 baseline model: forward flight (on
Figure 7.20: Location of the 60)hover
left) (and
the DC worst sector for
conditions theright).
(on the intake ]1 baseline model: forward
flight (on the left) and hover conditions (on the right).
For the forward flight condition, the maximum DC60 has a value of 0.45, and the worst
sector is located at 340 [deg]. On the other hand, in hover condition the maximum DC60 is
0.0205, and the worst sector is located at 360 [deg]. Moreover, a somehow more regular and
“periodical” fluctuation of distortions along the circumferential direction is evidenced in hover
(as can be appreciated in Figure 20), than that occurring in forward flight condition.
The initial assessment of the baseline intake aerodynamic features was essential for the
identification of a proper parameterization strategy. Actually, the intake optimization is
devoted to a reduction of the total pressure losses in both hover and forward flight conditions,
and also the maximum flow distortions at the AIP must be kept within acceptable values as
well. Hence, appropriate deformations will be carried out on the baseline geometry aimed at
reducing vortex flows and local separations inside the intake.
7.9 Parameterization 149

the baseline intake aerodynamic features was essential for the identification of a
proper parameterization strategy. Actually, the intake optimization is devoted to a
reduction of the total pressure losses in both hover and forward flight conditions,
and also the maximum flow distortions at the AIP must be kept within acceptable
values as well. Hence, appropriate deformations will be carried out on the baseline
geometry aimed at reducing vortex flows and local separations inside the intake.

7.9 Parameterization
Once the baseline CFD solution was analyzed, the geometry parameterization
was carried out. This operation is of outstanding importance in the optimization
process and it can be performed using various techniques. For the scope of the
present work, the commercial software Altair HyperMesh® was adopted as the
parameterization tool, due to the really versatile capabilities of the morphing tool
HyperMorph. The choice of this software is mainly justified with its effectiveness
and ease of use, which allows the user to build up complex parametric models in
a relatively easy way by means of a dedicated graphical user interface. Moreover,
the parameterization procedure in HyperMorph is very general: specifically, it is
independent from the peculiar model (either FEM or CFD) which is the object of
the optimization analysis. Therefore, whatever the geometry, the user is always
allowed to use the same morphing strategies, resulting in a remarkable saving of the
required working time for parameterization. Actually, both the above mentioned
characteristics are fundamental in an industrial context, where time is always an
essential issue. The geometry parameterization was carried out in a series of steps:

1. First of all, the baseline case file was imported into HyperMesh® in Fluent®
format.

2. The desired shapes were generated, starting from the baseline geometry,
through the morphing techniques available within HyperMorph®. For CFD
studies, the more suitable strategy is the domains-handles approach for
localized deformation, and the morph-volumes approach for global morphing
requirements.

3. The generated shapes were then saved: with this operation, HyperMorph
stores the current handles/nodes perturbations, allowing the user to apply
them to the baseline model with any given scaling factor. Using this approach,
the scaling factor of any generated shape can be dealt with as a design
variable by an optimization algorithm.

4. The parameterized model was saved into an .hm file, which was then used
for the batch-mode parameterization.

5. Finally, the HyperMesh® command file was created.


150 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

7.9.1 Morphing Constraints


HC/HIT09/WP1.2/D1/A (31/10/11)
Assessment of the optimization problem formulation and
application to a trial problem
The introduction of morphing constraints
(HEAVYcOPTer D1 partwasII) mandatory to obtain a suitable
optimized geometry for industrial applications. In fact, the requirement that the
final 8.1
intake design was
Morphing both able to be manufactured and installed into the aircraft
Constraints
led to the definition of a series of geometrical constraints (essentially related to
The introduction of morphing constraints was mandatory to obtain a suitable optimized
manufacturing and structural issues), that were considered in this preliminary
geometry for industrial applications. In fact, the requirement that the final intake design was
optimization.
both able to be manufactured and installed into the aircraft led to the definition of a series of
First of all, two directions were defined on the air intake surfaces, namely the
geometrical constraints (essentially related to manufacturing and structural issues), that were
“wetted” and “interior” directions: actually, the wetted direction is perpendicular
considered in this preliminary optimisation.
to the side ofall,the
First of twosurface that defined
directions were is wetted
on the by the intake
air intake surfaces, airflow
namely theand the interior
“"wetted"
direction is the directions:
and "interior" opposite, i.e. facing
actually, intodirection
the wetted the interior of the aircraft.
is perpendicular to the side of the
Referring to these two directions, the following geometric constraints
surface that is wetted by the intake airflow and the interior direction is the opposite, i.e. facingfor intake
optimization were
into the interior defined
of the aircraft. by AgustaWestland Ltd.:
Referring to these two directions, the following geometric constraints for intake
1. All the edges where the geometry attaches to the aircraft surface (AIP edges,
optimization were defined by AgustaWestland Ltd.:
topdeck edges) were a fixed constraint and could not be moved.
1. all the edges where the geometry attaches to the aircraft surface (AIP edges, topdeck
2. All the faces
edges) wereshown in redand
a fixed constraint in could
Figure 7.21
not be were not allowed to move in the
moved;
interior direction,
2. all the faces shownwhile
in redinin the wetted
Figure 22 weredirection there
not allowed wasin no
to move the constraint.
interior
direction, while in the wetted direction there was no constraint;
3. Finally, all the remaining faces, shown in green in Figure 7.21, were allowed
3. finally, all the remaining faces, shown in green in Figure 22, were allowed to move up
to move up to 30[mm] in the interior direction, while in the wetted direction
to 30 [mm] in the interior direction, while in the wetted direction there was no
thereconstraint.
was no constraint.

Free surfaces

Constrained surfaces

Figure 22: Geometrical constraints on the air intake surfaces: front view (on the left) and top
view (on the right).
Figure 7.21: Geometrical constraints on the air intake surfaces: front view (on the left) and
top view (on the right).
40 / 69 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
formal approval.

7.9.2 Design Variables Definition


In the present section, a brief description of the design parameters used for
the AW101 intake1 trial optimization is provided. A total amount of nine design
variables were considered. All the design variables were generated using the so
called domains-handles approach in order to satisfy global morphing requirements.
In particular, this approach allows for the application of mesh nodes displacements
7.9 Parameterization 151

within a geometrical region (domain) by changing the location of specific, user


defined, control points (handles) ([3]). Once applied, the nodes displacements may
be saved as perturbation vectors and then they can be re-applied to the baseline
model with any given scaling factor. Actually, the morphed geometry results from
a linear combination of the user defined shapes multiplied by their own scaling
factor:
9
v= ∑ αi Shi (7.4)
i =1
where:
• v is the global displacement vector;

• Shi is the ith basic shape;

• αi is the ith shape scaling factor and it is actually generated by the optimiza-
tion algorithm GeDEAII for each analyzed individual.
For the present application, the following ranges were defined for the values of
α
• αi ∈ [0, 1], i = 3, 4, 5, 6, 8, 9;

• αi ∈ [−1, 1], i = 1, 2, 7;
Using this approach, a scaling factor equal to zero means that the morphed
geometry is identical to the baseline one, while a scaling factor equal to one
produces the maximum allowed displacement within the specified range. Finally,
scaling factors equal to minus one produce the maximum allowed displacement
but in the opposite direction with respect to the original definition of the shape
modifications.
In the following, the adopted design variables for intake optimization are
described. In Figure 7.22, a series of cut planes used to define the shape functions
are illustrated, while the x-z or x-y sections of the intake model parametric shapes,
applied with a basic scaling factor equal to one, are reported from Figure 7.23 to
Figure 7.31; the corresponding handle displacements are visualized as well.
The selected shape functions are described hereafter:
1. sh1: this shape consists in an initial deformation along the z-axis of the blue
handle illustrated in Figure 7.23, which is located on the transversal cut
plane B; in this figure, the baseline shape along with the deformed one are
presented. The initial deformation range assigned to this variable was equal
to ± 20 [mm];

2. sh2: this shape consists in an initial deformation along the z-axis of the blue
handle illustrated in Figure 7.24, which lies on the transversal cut plane C.
The initial deformation range was given a value of ± 20 [mm].

3. sh3: this shape consists in an initial deformation along the z-axis of the blue
handle illustrated in Figure 7.62, located on the transversal cut plane B. It
was given an initial deformation range of ± 20 [mm];
152 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

Plane E

Plane D

Plane C

Plane B

Plane A

Figure 7.22: Cut planes used for the shape parameterization.

4. sh4:this shape consists in an initial deformation along the z-axis of the blue
handle illustrated in Figure 7.63, which lies on the transversal cut plane C. It
was given an initial deformation range of ± 20 [mm];

5. sh5:this shape consists in the deformation of the sharp corner located in the
upper portion of the air intake, close to the AIP, as illustrated in Figure 7.64.
It was given an initial deformation range of ± 20 [mm]; moreover, morph
constraints allowed to keep unchanged the red surfaces illustrated in Figure
7.64.

6. sh6:his shape consists in an initial deformation along the y-axis of the blue
handle located in the central region of the intake on the longitudinal cut
section E, as illustrated in Figure 7.28. It was given an initial deformation
range of ± 20 [mm] along the y direction;

7. sh7:this shape allows deforming the elbow region of the air intake both along
the x and y-axes, as depicted in Figure 7.29. The initial deformation range
assigned to this variable was equal to ± 20 [mm];

8. sh8:this shape consists in an initial deformation along both the x and y-axis of
the blue handle located in the central region of the intake on the longitudinal
cut section E, as illustrated in Figure 7.30. It was given an initial deformation
range of +20 [mm] and -20 [mm] along the x and y direction respectively;

9. sh9:this shape consists in a clockwise rotation of the whole air intake, around
the base point called “A” in Figure 7.31. The initially selected deformation
range was equal to ± 10 [deg]. Only clockwise rotation was allowed, in order
to comply with the geometrical constraints illustrated in Figure 7.21.
7.9 Parameterization 800-91-067 153
Numerical optimization of the EH101 Rev. A
Pag. 49 of 86
800-91-067
Numerical optimization of the EH101 Rev. A

Sh1 Pag. 49 of 86
Handle displacement vector

x-z plane section


Baselineconfiguration
Baseline configuration
Plane
Sh1 B Handle displacement vector
Deformed configuration
Δx= 0 [mm]
x-z plane section
Baselineconfiguration
Baseline configuration
Δy= 0 [mm]
Plane B Initial handle position

Deformed configuration
Δx= 0 [mm] Final handle position
Δy= 0 [mm] Initial handle position

Final handle position

Figure 7.23: Parametric shape shape


Figure 31 - Parametric Sh1, Sh1,
applied
appliedto
tothe intake
the intake model
model with factor
with scaling scaling factor α = +1.
α1=+1.

Sh2 Handle displacement vector


Figure 31 - Parametric shape Sh1, applied to the intake model with scaling factor α1=+1.
x-z plane section Baseline configuration

Plane
Sh2 C Handle displacement vector

Δx= 0 [mm]
x-z plane section
Deformed configuration
Baseline configuration
Δy= 0 [mm]
Plane C Initial handle position

Δy= 0
Δx= -20 [mm]
[mm] Deformed configuration
Final handle position

Δy= 0 [mm] Initial handle position

Δy= -20 [mm] Final handle position

Figure 32 - Parametric shape Sh2, applied to the intake model with scaling factor α1=+1.

Figure 32 - Parametric shape Sh2, applied to the intake model with scaling factor α800-91-067
1=+1.
Figure 7.24: Parametric shape Sh2, optimization
Numerical applied to the intake
of the model with scaling
EH101 Rev. A factor α = +1.
Pag. 50 of 86
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
Handle displacement vector
Sh3
Copyright – 2007
This document contains Agusta
confidential SpA,
proprietary Westland
information Helicopters
and is supplied Ltd, Westland
on the express condition that it may not beTransmissions Ltd orand
disclosed, reproduced In whole in part,AgustaWestland
or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
x-z plane section
International Ltd. (Collectively known as “AgustaWestland”) Baseline configuration

Copyright
Plane B and all other rights in this document are vested in AgustaWestland.
Deformed configuration
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose

Δx= 0 [mm]
other than for which it is supplied, without the written consent of AgustaWestland.

Initial handle position


Δy= 0 [mm]
Final handle position
Δy= 20 [mm]

Figure 7.25: Parametric


Figure shape
33 - Parametric Sh3,
shape Sh3,applied
applied toto
thethe intake
intake modelmodel withfactor
with scaling scaling
α1=+1.factor α = +1.

Sh4 Handle displacement vector

x-z plane section Baseline configuration

Plane C
Deformed configuration
Δx= 0 [mm]
Initial handle position
Δy= 0 [mm]
Δy= 20 [mm] Final handle position
154 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

Figure 33 - Parametric shape Sh3, applied to the intake model with scaling factor α1=+1.

Sh4 Handle displacement vector

x-z plane section Baseline configuration

Plane C
Deformed configuration
Δx= 0 [mm]
Initial handle position
Δy= 0 [mm]
Δy= 20 [mm] Final handle position

Figure 34 - Parametric shape Sh4, applied to the intake model with scaling factor α1=+1.
Figure 7.26: Parametric shape Sh4, applied to the intake model with scaling factor α = +1.
800-91-067
Numerical optimization of the EH101 Rev. A
800-91-067
Pag. 51 of 86
Numerical optimization of the EH101 Rev. A
Pag. 51 of 86

Copyright – 2007 Agusta SpA, Westland Helicopters


Handle displacement vector Ltd, Westland Transmissions Ltd and AgustaWestland
Sh5
International Ltd. (Collectively knownBaseline
as “AgustaWestland”)
configuration
Copyright x-z andplane section
all other rights in this document are vested
Handle displacement vector in AgustaWestland.
Sh5
This document Plane
contains confidential
D proprietary information and is supplied
Deformed on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
configuration
other than for which it is supplied, without the written consent of AgustaWestland.
Baseline configuration
x-zΔx= -20 [mm]
plane section Initial handle position

Δy= 0D[mm]
Plane Deformed
Final configuration
handle position
Δy= 0 [mm]
Δx= -20 [mm] Initial handle position

Δy= 0 [mm] Final handle position


Δy= 0 [mm]

Figure 35 - Parametric shape Sh5, applied to the intake model with scaling factor α1=+1.

Figure 7.27: Parametric shape Sh5, applied to the intake model with scaling factor α = +1.
Handle displacement vector
Sh6
Figure 35 - Parametric shape Sh5, applied to the intake model with scaling factor α1=+1.
Baseline configuration
x-y plane section
Plane E Deformed configuration

Δx= 15 [mm] Initial handle position


Handle displacement vector
Sh6
Δy= -15 [mm]
Baseline
Final configuration
handle position
Δy=
x-y 0 [mm]
plane section
Plane E Deformed configuration

Δx= 15 [mm] Initial handle position

Δy= -15 [mm]


Final handle position
Δy= 0 [mm]

Figure 36 - Parametric shape Sh6, applied to the intake model with scaling factor α1=+1.

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

Figure 36 - Parametric shape Sh6, applied to the intake model with scaling factor α1=+1.
Figure 7.28: Parametric shape Sh6, applied to the intake model with scaling factor α = +1.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
7.9 Parameterization 800-91-067 155
Numerical optimization of the EH101 Rev. A
Pag. 52 of 86
800-91-067
Numerical optimization of the EH101 Rev. A
Handle displacement vector
Sh7 Pag. 52 of 86
Baseline configuration
x-y plane section
Handle displacement vector
Sh7
Plane E Deformed configuration
Baseline configuration
Δx=plane
x-y section
20 [mm]
Initial handle
Initialposition
handle position
Δy= -20
Plane E [mm] Deformed configuration
Final handle position
Δx=
Δy= 0
20[mm]
[mm]
Initial handle
Initialposition
handle position
Δy= -20 [mm]
Final handle position
Δy= 0 [mm]

Figure 37 - Parametric shape Sh7, applied to the intake model with scaling factor α =+1.
Figure 7.29: Parametric shape Sh7, applied to the intake model with scaling
1
factor α = +1.
Figure
Sh8 37 - Parametric shape Sh7, applied to the intake model with scaling factor α1=+1.
Handle displacement vector

x-y plane section Baseline configuration


Handle displacement vector
Sh8
Plane E and B
Deformed configuration
Δx=plane
x-y section
20 [mm] Baseline configuration
Initial handle position
Δy= -20
Plane E [mm]
and B
Deformed configuration
Δx=
Δy= 0
20[mm]
[mm] Final handle position
Initial handle position
Δy= -20 [mm]
Δy= 0 [mm] Final handle position

800-91-067
Numerical optimization of the EH101 Rev. A
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
Pag. 53 of 86
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

Figure 7.30: Parametric shape Sh8, applied to the intake model with scaling factor α = +1.
Figure 38 - Parametric shape Sh8, applied to the intake model with scaling factor α1=+1.

Sh9 Handle displacement vector A


x-y plane section Baseline configuration
Plane A
Deformed configuration
Clockwise
rotation of 10 Initial handle position
[deg].
“A” is the base Final handle position

point of the
156 Part I 38
Figure - Aerodynamic shape
- Parametric shape Sh8, appliedoptimization of scaling
to the intake model with the AW101 air intake ]1
factor α =+1. 1

Sh9 Handle displacement vector A


x-y plane section Baseline configuration
Plane A
Deformed configuration
Clockwise
rotation of 10 Initial handle position
[deg].
“A” is the base Final handle position

point of the
rotation

Figure 39 - Parametric shape Sh9, applied to the intake model with scaling factor α1=+1.
Figure 7.31: Parametric shape Sh9, applied to the intake model with scaling factor α = +1.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland

7.10 Formulation of the intake optimization problem


International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
Once the design CFD model and the parametric model for the intake geometry
were built up, the successive step consisted in the GeDEAII-driven optimization.
As stated above, the optimization problem considered was actually of the multi-
point type, since the total pressure losses occurring in both forward flight and
hover conditions needed to be minimized simultaneously. Moreover, the maximum
DC (60) at AIP was required to be kept to reasonable levels in both the considered
flight conditions. From a mathematical point of view, the optimization problem
can be expressed in the following way:

Minimize [ F (x)] (7.5)

where [ F (x)] is a vector function:


(
∆Ptot | @ f orward f light
[ F (x)] = (7.6)
∆Ptot | @ hovering f light

and x is the design variable vector, consisting in the coefficients of the linear
combination of the shape functions describing the morphed intake geometries
(please read section 7.9):

x = [ α1 , α2 , α3 , α4 , α5 , α6 , α7 , α8 , α1 9] (7.7)

subject to the following variables bounds:

• αi ∈ [0, 1], i = 3, 4, 5, 6, 8, 9;

• αi ∈ [−1, 1], i = 1, 2, 7;

Thanks to its multi-objective formulation, the optimization algorithm seeks for


solutions featuring improved performance in terms of total pressure loss for both
the flight conditions at hand. As mentioned in section 6.4.1, the mass-weighted
7.11 Discussion of results 157

average integral of the total pressure over the AIP was used in the present work for
computation of total pressure losses. The requirement of keeping the maximum
DC (60) equal or lower than the baseline value in both the flight conditions was
treated as a functional constraint, and a penalty function approach was used to
handle this constraint. Specifically, some additional terms were added to the fitness
functions expressed in Equation 7.6, which penalized the fitness values when a
constraint was violated. In fact, a constraint violation results in an increase of the
additional term which in turn increases the original value of the fitness function.
As a consequence, during the selection mechanism the evolutionary algorithm
tends to prefer individuals that do not violate the constraints, and to discard the
penalized ones. In this way, it is likely that the latest generation will be constituted
by individuals that do not violate these constraints, and whose penalty function
values have consequently null value. The adopted penalty function for DC (60)
coefficient takes the following form:

0 i f DC60con f iguration ≤ DC60baseline
PF = α (7.8)
 β | DC60con f iguration − DC60baseline |

DC60baseline i f DC60con f iguration > DC60baseline

where α is equal to 0.5, in order to give a convex form to the penalty function
and force the constraints satisfaction; β was given a value equal to 700 and 100
for forward flight and hovering conditions respectively. These specific values were
selected after a preliminary analysis aimed at finding a suitable shape for the
penalty functions curves. The choice of the absolute value of total pressure losses
rather than the total pressure ratio for the fitness function evaluation was actually
related to the use of the above mentioned penalty function for control of the
maximum flow distortion. In fact, being the values of the pressure recovery very
close to 1, a suitable coupling with the typical values of the penalty function would
have been more difficult, thus making the penalization for constraint violation
ineffective. Therefore, the optimization problem can be finally reformulated as
follows:
Minimize [ PF + F (x)] (7.9)
Moreover, the number of individuals per generation was set to 12, while the
number of generations for the trial optimization run was set to 20.

7.11 Discussion of results


The trial intake optimization results are discussed in this section. First of all, in
Figure 7.32, the overall set of simulated individuals is represented in the objective
space. In particular, results refer to the modified fitness function specified in
Equation 7.9, in which the penalty functions are taken into account as well. It is
worth recalling that the numerical total pressure losses were determined using a
mass-weighted average surface integral of the total pressure over the AIP surface.
Finally, the ultimate Pareto front after 20 generations is illustrated in Figure
7.33. Also in this case, the modified objective functions with addition of the penalty
First of all, in Figure 33, the overall set of simulated individuals is represented in the
objective space. In particular, results refer to the modified fitness function specified in Eq. 11,
in which the penalty functions are taken into account as well. It is worth recalling that the
158
numerical totalPart I - Aerodynamic
pressure shape optimization
losses were determined of the AW101
using a mass-weighted air intake
average ]1
surface
integral of the total pressure over the AIP surface.

Figure 33: The entire set of geometries calculated during the optimization run.
Figure 7.32: The entire set of geometries calculated during the optimization run.
49 / 61 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
formal approval.
functions are represented. In addition, the DC (60) values for each individual
on the Pareto front are reported for both forward flight and hover conditions.
From the figure, the conflicting nature of the two selected objectives is clearly
apparent. Furthermore, the front appears to be very sharp, in the sense that very
small improvements of the total pressure losses in forward flight condition lead
to remarkable degradations of the corresponding objective function in hover. In
addition, neither the DC (60) in forward flight nor in hover conditions show a
monotonic trend with the values of the objective functions along the Pareto front.
The point on the Pareto front represented with a star in Figure 7.33 was selected
as the final solution, since it was judged a satisfactory compromise between the two
selected objectives. In Table 7.12, the design variables values of the selected solution
are summarized (a null value indicates that the corresponding shape function was
not modified with respect to the baseline). Moreover, a direct comparison of the
optimized solution geometry with the baseline is illustrated in Figure 7.34. As

Parameter ] Sh1 Sh2 Sh3 Sh4 Sh5 Sh6 Sh7 Sh8 Sh9
Selected solution −0.7 −0.373 0 -1 0 0 1 0.1686 1

Table 7.12: Trial intake optimization results: design parameters values for the selected
optimized individual.

apparent from Figure 7.34, the optimized solution features an enlarged passage
area at the transversal mid-section (sh1), which allows a local rearrangement of
(HEAVYcOPTer D1 part II)

Moreover, in Figure 34, the evolution of the Pareto front throughout the whole
optimization is depicted:
7.11 Discussion the fitness functions improvement with increasing number 159
of results of
generations is clearly appreciable.
Finally, the ultimate Pareto front after 20 generations is illustrated in Figure 35. Also in
this case, the modified objective functions with addition of the penalty functions are
represented. In addition, the DC60 values for each individual on the Pareto front are reported
for both forward flight and hover conditions. From the figure, the conflicting nature of the two
selected objectives is clearly apparent. Furthermore, the front appears to be very sharp, in
the sense that very small improvements of the total pressure losses in forward flight condition
lead to remarkable degradations of the corresponding objective function in hover. In addition,
neither the DC60 in forward flight nor in hover conditions show a monotonic trend with the
values of the objective functions along the Pareto front.

Figure 35: Final Pareto front after 20 generations.


Figure 7.33: Final Pareto front after 20 generations.

51 / 61 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
formal approval.
Parameter# Sh1 Sh2 Sh3 Sh4 Sh5 Sh6 Sh7 Sh8 Sh9
Selected
-0.7 -0.373 0 -1 0 0 1 0.1686 1
solution
160 Part I - Aerodynamic shape optimization of the AW101 air intake ]1
Table 12: Trial intake optimization results: design parameters values for the selected optimised
individual.

Baseline
Optimized

Figure 36: Optimized solution geometrical configuration: comparison with the baseline intake
geometry.
Figure 7.34: Optimized solution geometrical configuration: comparison with the baseline
intake geometry.
52 / 61 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
formal approval.

the flow field, and therefore a reduction of the vortex flow, as will be illustrated
later on. On the other hand, deformation of the sh4 shape acts in the direction
of locally widen the intake duct section, and this has beneficial effects on both
total pressure losses reduction and flow distortion requirements. Moreover, the
optimized solution features a more smoothed elbow curvature (sh7) with respect
to baseline. Finally, a remarkable rotation of the whole air intake around the point
“A” depicted in Figure 7.34 is given to the optimized configuration: once again,
this has a beneficial impact on the flow field inside the intake duct, since it reduces
flow separation in the entry region occurring in forward flight condition. In Table
7.13, the values of the objective functions of the optimized solution are reported
and compared with the baseline. As apparent, a remarkable reduction of the total
pressure losses along the intake duct was obtained with the optimized solution,
especially in hover conditions; in addition, regarding the flow distortion at the AIP,
a large reduction of the DC (60) at the engine face was achieved in forward flight
and, to a lesser extent, in hover conditions, even though the flow distortion was
not directly included in the objective function, but rather treated using a penalty
function (please read section 7.32).
7.11 Discussion of results 161

Baseline configuration Forward Flight Hovering Flight


Total pressure drop reduction with respect to baseline [%] \ 10.85% 23.4%
DC60 reduction with respect to baseline [%] \ 19.6% 8.29%

Table 7.13: Trial intake optimization results: Total pressure drop and DC60 reduction with
respect to the baseline geometry

In Figure 7.35 and Figure 7.36 the total pressure contours over a series of
transversal sections along the optimized intake duct are depicted for both the
hovering and forward flight conditions and they are compared with the baseline.
HC/HIT09/WP1.2/D1/A (31/10/11)
As apparent, in hover conditions the region of low total pressure occurring in the
Assessment of the optimization problem formulation and
final portion of the intake duct is less extended
application to a trial in the optimized geometry, and
problem
total pressure losses are less severe. Furthermore,
(HEAVYcOPTer D1 also
part in
II) forward flight conditions

Baseline Optimised

Reduction of total
pressure losses

Figure 37: Contours of total pressure ([Pa]) over a series of transversal sections along the
Figure 7.35:
intake Contours
duct in hoverofconditions:
total pressure ([Pa]) over
comparison a series
of the of transversal
baseline (on the left)sections along the
and optimized (on the
intake duct in hovering condition: comparisonright) solutions.
of the baseline (on the left) and optimized
(on the right) solutions.

Baseline Optimised
a reduction of the total pressure losses with respect to the baseline is evidenced, Reduction of total
pressure losses
especially in the upper portion of the intake duct and towards the AIP. As expected,
also the flow behavior over the AIP is improved in the optimized solution with
respect to the baseline, for both hovering and forward flight conditions, as apparent
from Figure 7.37 and Figure 7.38: in fact, the total pressure field is much more
uniform over the engine face in both the considered flight conditions, and the
most severe total pressure drops are eliminated in the optimized geometry. The
more favorable behavior of the optimized intake duct is confirmed also by the
visualization of the streamlines path, illustrated in Figure 7.39 and Figure 7.40
for hovering and forward flight conditions respectively. While no appreciable
differences are evidenced in the streamlines behavior with respect to the baseline
in hovering, a remarkable reduction of the vortex flow occurring in the second
Figure 38: Contours of total pressure ([Pa]) over a series of transversal sections along the
intake duct in forward flight conditions: comparison of the baseline (on the left) and optimized
(on the right) solutions.

Furthermore, also in forward flight conditions a reduction of the total pressure losses with
respect to the baseline is evidenced, especially in the upper portion of the intake duct and
towards the AIP.
As expected, also the flow behavior over the AIP is improved in the optimized solution
162Figure 37: Contours
Part of total pressure
I - Aerodynamic ([Pa])
shape over a series of
optimization of transversal
the AW101sections along
air intake ]1the
intake duct in hover conditions: comparison of the baseline (on the left) and optimized (on the
right) solutions.

Baseline Optimised
Reduction of total
pressure losses

Figure 7.36: Contours of total pressure ([Pa]) over a series of transversal sections along
Figure 38: Contours of total pressure ([Pa]) over a series of transversal sections along the
the intake
intake ductduct in forward
in forward flightflight condition:
conditions: comparison
comparison of theofbaseline
the baseline (onleft)
(on the theand
left)optimized
and
optimized (on the right) solutions. (on the right) solutions.

Furthermore, also in forward flight conditions a reduction of the total pressure losses with
bend of the S-shaped duct is achieved in forward flight conditions, due to the
respect to the baseline is evidenced, especially in the upper portion of the intake duct and
more streamlined shape of the optimized geometry with respect to the baseline.
towards the AIP.
Actually, this beneficial effect is propagated to the AIP, as illustrated in Figure 7.38.
As expected,
Finally, also the of
the comparisons flowthebehavior
baselineover andthe AIP is improved
optimized DC60 profilesin the over
optimized solution
the AIP
for
withboth hovering
respect and forward
to the baseline, for both flight
hover conditions
and forward are reported
flight in Figure
conditions, 7.41 and
as apparent from
Figure 7.42 respectively. While no appreciable modifications are shown in hover, a
remarkable reduction of the maximum DC60 value is apparent in forward flight
54 / 61 This document
conditions with respect to the
is the property baseline
of the author(s) (see Table 7.12):
organization(s) this
and shall is distributed
not be in fact very beneficial
or reproduced without their
formal approval.
from the compressor performance point of view.

7.12 Reverse Engineering Procedure in CATIA® V5


From the optimization run the values of the design variables of the selected
geometry were derived. Those values were introduced within Altair Hypermesh®
in order to obtain the morphed mesh of the optimal geometry. Then, the mesh was
exported in stl format: specifically, only the modified components were exported,
being the dummy duct left unchanged.
The reverse engineering process results straightforward and it is qualitatively
described in the following.
First, the file *.stl created in Altair Hypermesh® was read within CATIA® V5
using of the Import function under the Digitalize Shape Editor module.
Second, the Quick Surface Reconstruction module was used in order reconstruct
the surfaces from the stl model. In particular, a value of 1000 for Surface Details
option was selected, and the Auto Tangency option was activated before launching
the reconstruction process. This allowed to properly reconstruct the surfaces, and
HC/HIT09/WP1.2/D1/A (31/10/11)
Assessment of the optimization problem formulation and
application to a trial problem
(HEAVYcOPTer D1 part II)
7.12 Reverse Engineering Procedure in CATIA® V5 163
Figure 39 and Figure 40: in fact, the total pressure field is much more uniform over the
engine face in both the considered flight conditions, and the most severe total pressure drops
HC/HIT09/WP1.2/D1/A (31/10/11)
are eliminatedAssessment
in the optimized geometry.
of the optimization problem formulation and
application to a trial problem
(HEAVYcOPTer D1 part II)
Baseline Optimised

Figure 39 and Figure 40: in fact, the total pressure field is much more uniform over the
Reduction of total
engine face in both the considered flight conditions, and the most severe total pressure pressure
drops losses
are eliminated in the optimized geometry.

Baseline Optimised

Reduction of total
pressure losses

Figure 7.37: Contours of total pressure ([Pa]) over a series of transversal sections along the
Figure 39: Total pressure contours ([Pa]) over the AIP in hover: comparison of baseline (on the
intake duct in hovering condition: comparison
left) and of the
optimized (on the baseline (on the left) and optimized
right) solutions.
(on the right) solutions.
Baseline Optimised

Reduction of total
pressure losses

Figure 39: Total pressure contours ([Pa]) over the AIP in hover: comparison of baseline (on the
left) and optimized (on the right) solutions.

Baseline Optimised

Reduction of total
pressure losses

Figure 40: Total pressure contours ([Pa]) over the AIP in forward flight: comparison of baseline
(on the left) and optimized (on the right) solutions.

The more favorable behavior of the optimized intake duct is confirmed also by the
visualization of the streamlines path, illustrated in Figure 41 and Figure 42 for hover and
forward flight conditions respectively. While no appreciable differences are evidenced in the
streamlines behavior with respect to the baseline in hovering, a remarkable reduction of the
Figure
vortex40: Total
flow pressure
occurring in contours
the second ([Pa])
bendoverof the
theAIP in forward
S-shaped ductflight: comparison
is achieved of baseline
in forward flight
Figure 7.38: Contours of (ontotal
the pressure ([Pa]) over
left) and optimized (ona the
series of transversal
right) solutions. sections along
the55 / 61 This
intake duct in forward
document flightofconditions:
is the property comparison
the author(s) organization(s) andof the
shall notbaseline (onor the
be distributed left) and
reproduced without their
formal approval.
The more
optimized (on thefavorable behavior of the optimized intake duct is confirmed also by the
right) solutions.
visualization of the streamlines path, illustrated in Figure 41 and Figure 42 for hover and
forward flight conditions respectively. While no appreciable differences are evidenced in the
streamlines behavior with respect to the baseline in hovering, a remarkable reduction of the
vortex flow occurring in the second bend of the S-shaped duct is achieved in forward flight
55 / 61 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
formal approval.
AssessmentHC/HIT09/WP1.2/D1/A
of the optimization problem formulation and
(31/10/11)
application to a trial problem
Assessment of the optimization problem formulation and
(HEAVYcOPTer
application D1 problem
to a trial part II)
(HEAVYcOPTer D1 part II)
conditions, due to the more streamlined shape of the optimized geometry with respect to the
164conditions,
Part
due Ito the
baseline. Actually, -this
Aerodynamic
more effectshape
streamlined
beneficial isshape optimization
of the to
propagated optimized of the AW101
the AIP, geometry air intake
withinrespect
as illustrated Figure 40.the ]1
to
baseline. Actually, this beneficial effect is propagated to the AIP, as illustrated in Figure 40.

Baseline Optimised
Baseline Optimised

Figure 41:Streamlines
7.39:
Figure pattern
Streamlines pattern inside
inside the the air intake
air intake in comparison
in hover: hover: comparison
of baselineof(on
baseline
the (on
the left) and optimized left)
(on and
theoptimised
right) (on the right) configurations.
configurations.
Figure 41: Streamlines pattern inside the air intake in hover: comparison of baseline (on the
left) and optimised (on the right) configurations.
Baseline Optimised
Baseline Optimised

HC/HIT09/WP1.2/D1/A (31/10/11) Flow vorticity


Assessment of the optimization problem formulation and reduction
Flow vorticity
application to a trial problem reduction
(HEAVYcOPTer D1 part II)

maximum DC60
Figure value is apparent
42: Streamlines in forward
pattern inside flightinconditions
the air intake forward flight:with respect
comparison to the baseline
of baseline
(on the left) and optimised (on the right) configurations.
Figure 42: Streamlines
7.40:
Figure pattern inside
the airthe airinintake
forwardin forward flight:ofcomparison of
(see Table 13): this isStreamlines
in fact pattern
very inside
beneficial from intake
the flight:
compressor comparison
performance
(on the left) and optimised (on the right) configurations.
baseline
point of view.
baseline (on the left) and optimized (on the right) configurations.
Finally, the comparisons of the baseline and optimized DC60 profiles over the AIP for
bothFinally, the forward
hover and comparisons of the baseline
flight conditions and optimized
are reported in FigureDC60 profiles
43 and Figureover the AIP for
44 respectively.
both
Whilehover and forward modifications
no appreciable flight conditions areshown
are reported
in in Figurea 43
hover, and Figurereduction
remarkable 44 respectively.
of the
56 / 61 no
While This appreciable
document is the property of the author(s)
modifications are organization(s)
shown in and shall not
hover, a beremarkable
distributed or reproduced
reduction without
of their
the
formal approval.
56 / 61 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
formal approval.

Figure 43: Comparison of baseline and


Figure 7.41: Comparison optimised
of baseline and DC60 distribution
optimized in hover. in hovering.
DC60 distribution
7.12 Reverse Engineering Procedure in CATIA® V5 165
Figure 43: Comparison of baseline and optimised DC60 distribution in hover.

Figure 44: Comparison


Figure of baseline
7.42: Comparison and optimised
of baseline DC60DC60
and optimized distribution in forward
distribution flight.flight.
in forward

to capture all the needed details.


The upper and isometric views of the optimized geometry are depicted in
Figure 7.43.
Once the optimized model has been reconstructed, in order to evaluate the
geometrical differences between the optimized and the baseline configurations, a
comparison was performed. It consisted in an overlap of the two CAD models.

The superimposition is depicted in Figure 7.44.

57 / 61 This document is the property of the author(s) organization(s) and shall not be distributed or reproduced without their
formal approval.
166 Part I - Aerodynamic shape optimization of the AW101 air intake ]1

(b) Iso

(a) Upper view (a) Upper view


(b) Isometric view

Figure 7.43: The final optimized intake geometry.


Figure 7.43: The final optimized intake geometry.

Figure 7.44: Superimposition of the baseline and the optimized intake CAD model.
Part II

Test case B:
Aerodynamic shape optimization
of the frontal region of the ERICA
tiltrotor using Ansys Fluent® as
the CFD solver

167
169

7.13 Introduction
The present work deals with the ERICA tiltrotor nose shape optimization
aimed at reducing its baseline aerodynamic drag. To this purpose, several steps
were performed in order to obtain reliable results while reducing the overall
optimization time as much as possible.
First, a study of drag sensitivity to the mesh quality was carried out (as
described in chapter 7.19) in order to find out a model capable of reproducing drag
characteristics of the tiltrotor nose with an acceptable accuracy while featuring the
smallest number of 3D elements. As a matter of fact, decreasing the number of grid
elements results in a reduction of calculation time, but, at the same time, the model
capacity to properly reproduce the aerodynamic behaviour of this component
is progressively deteriorated. Hence, the goal of this first step was to identify a
suitable model resulting from a trade-off between these two aspects.
The second step consisted in a D.O.E. analysis over the selected CFD model,
aimed at determining which factors were most influential on the response, that
is, which design variables most influenced the nose drag characteristics. To this
purpose, a Student analysis was carried out on the independent variables. The
shape parameterization, the D.O.E. analysis and the Student test are illustrated
and discussed in chapter 7.22.
Finally, the nose aerodynamic shape optimization was performed, as discussed
in detail in chapter 7.23. As regards the optimization engine, the GeDEAII algo-
rithm was selected, whose characteristics and performance were described during
chapter 2.

7.14 Optimization Procedure


In this chapter, the whole sequence of operations necessary to carry out the
optimization study is described. As already mentioned, the optimization procedure
was not performed in a single step, but rather multiple steps were required to
obtain reliable results, while reducing the overall optimization time as much as
possible. Each operation will be fully analyzed in the next chapters.
The whole sequence of operations carried out to address a multi-objective
optimization problem is shown in Figure 7.45. It is worth mentioning that each
block was carried out independently from the others.
The upper portion of the flow chart shows the set of preliminary operations
that need to be accomplished before starting the successive automatic loops. The
procedure started from the elaboration of a baseline geometry CAD model of the
tiltrotor fuselage provided by AgustaWestland in CATIA® format. The following
steps focused on the construction of the baseline CFD model: first of all, the surface
mesh was generated using CATIA® surface meshing tool, then the volume mesh
was created by means of Ansys TGrid® and finally the CFD case was set up with
the commercial CFD code Ansys Fluent® which is the standard code for CFD
simulation adopted at AgustaWestland.
170 Part II - ERICA nose shape optimization using Ansys Fluent®

Next, a study on aerodynamic drag sensitivity to the grid features was carried
out as described in detail in chapter 7.19. In fact, it is well known that cross-sectional
area dimensions, virtual wind tunnel box length and mesh refinement level greatly
influence the computed drag. A lot of combinations of these parameters were
tested to obtain a model characterized by an acceptable accuracy while featuring
the least possible number of cells.
Once this study was performed, the selected model was “frozen” and imported
into Altair HyperMesh® environment for the geometrical parameterization.
A D.O.E. study was then carried out with the aim of exploring the design space,
and a Student analysis was undertaken which made it possible to decrease the
number of independent design variables; hence, the total time needed for the next
GeDEAII driven optimization procedure could be reduced.
Once the independent variables to be retained were identified, the optimization
process could be addressed, which was managed by GeDEAII as the master
program.
Here some words are spent about the followed morphing strategy in this
work. In details, it was decided to parameterize the surface mesh only, and to
re-mesh every time the volume mesh, for the reasons explained below. Actually,
the direct volume mesh morphing may cause boundary layer prisms distortions,
leading to a deformed mesh featuring elements having negative Jacobian, even for
small displacements. Instead, this choice allowed us to overcome this morphing
limitation: it is now possible to morph only the surface mesh and then re-mesh
the volume at each objective function/s evaluation, with the only drawback of a
minimal increased overall optimization time.
Moreover, the surface mesh morphing, and then the volume re-meshing, al-
lowed to increase the morphing capabilities and therefore to enlarge the optimiza-
tion search space.
The lower part of the flow chart in Figure 7.45 describes the results extraction
and elaboration. The output of the automatic optimization loop is the optimal
combination of the design variables of the individuals lying on the Pareto frontier.
At this point, the designer can identify the individuals of interest among the
multiple optimal solutions of the Pareto front and the meshed geometry of each
selected individual can be reconstructed using HyperMorph®.
The final step is the reverse engineering process, which allows obtaining the
deformed surface starting from the mesh; this process can be performed using
CATIA®.

7.15 The object of the optimization


The scope of this work consisted in the optimization of ERICA nose, whose
1/8th scaled wind-tunnel model is depicted in Figure 7.46. To this purpose, a
CFD-based optimization procedure was implemented, which was introduced in
chapter 7.14, and will be described in detail in the following chapters.
CFD simulations were carried out on the components depicted in Figure 7.47. In
7.15 The object of the optimization 171

Sensitivity
Geometry elaboration anaysis
CATIA & Tgrid Volume mesh
surface mesh

FLUENT

Parameterization
HyperMesh
Optimal mesh
model
& Parameterization
HyperMorph

D.O.E. +
Student
D.O.E. Design variables Morphed surface analysis
Hypermesh
Analysis values mesh

CFD objective
function values
FLUENT Volume mesh Tgrid

Student Reduction of the


Analysis design variables

GeDEAII
Design variables Morphed surface driven
GeDEAII values
Hypermesh
mesh
optimization

CFD objective
function values
FLUENT Volume mesh Tgrid

Pareto front
optimal design
variables
combination

HyperMesh Optimal
Optimal
& individuals mesh CATIA geometries
HyperMorph reconstruction

Figure 7.45: Flow chart of the complete optimization procedure.

addition to these components, a tail cone (Figure 7.48) was added to the simulated
geometry in order to avoid flow separations downstream the cylindrical portion of
the fuselage which could detrimentally affect convergence behavior of numerical
simulations.

As far as the optimization target is concerned, a reduction of 2% in the Cd of the


baseline ERICA nose was required. The reference flight conditions are summarized
in Table 7.14.

The reader is referred to the Acronyms chapter for the definition of fuselage
incidence.
7. The object of the optimization
The scope of this work consisted in the optimization of ERICA nose, whose 1/8th scaled wind-
tunnel model is172 Part II - ERICA nose shape optimization using Ansys Fluent®
depicted in Figure 2. 800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 16 of 98

7. The object of the optimization


The scope of this work consisted in the optimization of ERICA nose, whose 1/8th scaled wind-
tunnel model is depicted in Figure 2.

Figure 2: ERICA nose geometry [7].


 
To this purpose, a CFD-based optimization procedure was implemented, which was introduced

Figure
Figure 2: ERICA
7.46: ERICA nosenose geometry
geometry [7]. [74])
(taken from
in chapter 5, and will be described in detail in the following chapters.
CFD simulations were carried out on the components depicted in Figure 3.

To this purpose, a CFD-based optimization procedure was implemented, which was introduced
in chapter 5, and will be described in detail in the following chapters.
CFD simulations were carried out on the components depicted in Figure 3. 800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 17 of 98

   
In addition to these components, a tail cone (Figure 4) was added to the simulated geometry in
Figureflow
order to avoid ERICA nose downstream
7.47:separations
Figure 3: ERICA (on the right) the
nose (on the right) and and cylindrical
cylindrical portion
cylindrical portion of
portion of
the fuselage
of the
(on the
thefuselage
left) [7]. (onwhich
fuselage the left)
could
(taken from [74])
detrimentally affectCopyright
convergence behaviour of numerical simulations.
– 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

   

Figure 3: ERICA nose (on the right) and cylindrical portion of the fuselage (on the left) [7].
 

Figure
Figure 4: Tail cone to ERICA
7.48:
added tail cone
the fuselage (taken from
components for[74])
CFD simulations.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
As far as the optimization target is concerned, a reduction of 2% in the Cd of the baseline
Copyright and all other rights in this document are vested in AgustaWestland.
ERICA
This document nose
contains was required.
confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
The reference flight conditions are summarized in Table 1.
7.16 Geometry elaboration 173

Flight True True Pressure Static OAT Density Speed Fuselage Nacelle Outer Rotor One
Condition air air alti- pres- [C ] [kg/m3 ] of inci- atti- wing RPM rotor
speed, speed, tude, sure, sound dence tude setting thrust,
[Kts] [m/s] [m] [ Pa] [m/s] [deg] [deg] angle [N]
[deg]
Forward 350 180.06 7500 38251 -33.75 0.556623 310.175 -1.97 0.00 0.00 425.8 12431
Flight cruise

Table 7.14: Flight Conditions for Nose optimization (taken from [74])

7.16 Geometry elaboration


As already mentioned, the tiltrotor model was made up of three components,
specifically the canopy, the fuselage and the tail cone, depicted in Figure 7.47 and
Figure 7.48 respectively. The aircraft fuselage was inserted into a virtual wind
tunnel, whose dimensions were varied during the sensitivity analysis, as will be
described later. A sketch of the wind tunnel box is depicted in Figure 7.49.

Figure 7.49: Virtual wind-tunnel layout with the tiltrotor fuselage.

In CATIA® environment, the model surfaces were organized into six distinct
components, namely:

1. Inlet, which is depicted in Figure 7.49;

2. Outlet, which is depicted in Figure 7.49;

3. Symmetry, which is depicted in Figure 7.49;

4. Nose, which corresponds to the canopy surface, depicted in Figure 7.46;


174 Part II - ERICA nose shape optimization using Ansys Fluent®

5. Fuselage, which corresponds to the cylindrical portion of the fuselage surface,


depicted in Figure 7.47;

6. Tail cone, which corresponds to the tail cone surface, depicted in Figure 7.48.

7.17 Mesh generation


As far as the mesh generation is concerned, the software used to carry out the
meshing operations was respectively:

1. CATIA® surface meshing tool for the surface mesh generation;

2. Ansys Tgrid® for the volume mesh generation.

7.17.1 Superficial mesh


The CATIA® surface meshing tool can be used to mesh the geometry surfaces
directly from the CATIA® format, so no import/export operations are necessary
to accomplish this task. Moreover, the surface mesh obtained from the CATIA®
meshing tool is entirely interfaced with the CATIA® geometry. This characteristic is
important in an industrial context, because a modification of the baseline geometry
can be directly transformed into a correspondent surface mesh modification simply
by updating the mesh.
The surface mesh was more refined over the tiltrotor nose, which was the object
of optimization, whereas the superficial grid over the cylindrical portion of the
fuselage was slightly coarser; finally, the tail cone surface mesh was the coarsest
one. The choice of adopting different mesh sizes was motivated by the attempt to
contain the total amount of surface elements (and consequently the number of 3D
cells) as much as possible.
The transition between two zones with different mesh size was accomplished
by properly selecting CATIA® Advanced meshing tool parameters. This feature
is a really important characteristic of a CFD mesh, and it is fundamental for
a good CFD solution convergence. Two different surface mesh distributions on
the ERICA surface were investigated during the sensitivity analysis, as will be
illustrated in chapter 7.19. The finally selected model was characterized by a
number of triangular elements equal to 175.191; the corresponding target element
size (expressed in [m]) and the number of superficial elements for each of the
fuselage components are summarized in Table 7.15.
A sketch of the finally selected surface mesh over the nose component is
reported in Figure 7.50, while the superficial mesh over the virtual wind tunnel
walls is illustrated in Figure 7.51. Some words are worth spending as regards the
quality criteria accomplished during superficial mesh generation. Actually, in order
to achieve a good mesh quality it was necessary to define and set proper quality
parameters before starting the meshing operations; these parameters were then
checked during the mesh generation. AgustaWestland quality requirements can
7.17 Mesh generation 175

Number of elements Target element size [mm]


800-91-067
Inlet 1200 3500
Numerical optimisation of ERICA tiltrotor nose Rev. A
Outlet 1200 3500 Pag. 20 of 98
Symmetry 6512 3500
Canopy
The finally 73648
selected model was characterized 30
by a number of triangular elements equal to
175,191; theFuselage 76994size (expressed in mm) and40
corresponding target element the number of superficial
elements for Tail
eachcone 15636
of the fuselage components are summarized in Table 80
2.

Table 7.15: Finally selected number ofNumber


elements and target element
of elements Targetsize for size
element each fuselage
component
Inlet 1200 3500
Outlet 1200 3500
Symmetry 6512 3500
Canopy 73648 30
Fuselage 76994 40
Tail cone 15636 80
Table 2: Finally selected number of elements and target element size for each fuselage
component.

A sketch of the finally selected surface mesh over the nose component is reported in Figure 6,
while the superficial mesh over the virtual wind tunnel walls is illustrated in Figure 7.

Figure 6: Superficial mesh over ERICA nose.


Figure 7.50: Superficial mesh over ERICA nose.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 21 of 98

Some words are worth spending as regards the quality criteria accomplished during superficial
176 Part II - ERICA nose shape optimization using Ansys Fluent®
mesh generation. Actually, in order to achieve a good mesh quality it was necessary to define and
set proper quality parameters before starting the meshing operations; these parameters were then
be met by imposing
checked proper
during the mesh limitsAgustaWestland
generation. on Skewnessquality
and requirements
Aspect-Ratiocanwithin
be met CATIA®.
by
Theseimposing
parameters canonbe
proper limits defined
Skewness as follows:
and Aspect-Ratio within CATIA®.

Figure 7: Superficial mesh over the virtual wind tunnel walls.


Figure 7.51: Superficial mesh over the virtual wind tunnel walls.
These parameters can be defined as follows:

• Skewness: it is a quantitative measure of the element distortion with respect


• Skewness: it is a quantitative measure of the element distortion with respect to the ideal
to the element
ideal element shape (equilateral triangle for triangular surface mesh).
shape (equilateral triangle for triangular surface mesh).
is calculated
Skewness
Skewness is calculatedwithin CATIA®
within CATIA® in the way:
in the following following way:

S
QQ== 11−−S (7.10)
S′ S0
where Q is the Skewness, S is the ideal element surface, and S’ is the actual element
where Q is the Skewness, S is the ideal element surface, and S0 is the actual
surface.
element surface.

• Aspect-Ratio: it is the ratio between the maximum and minimum length


among
Copyright – the
2007 sides of Westland
Agusta SpA, the element;
Helicopters for instance,
Ltd, Westland the Aspect-Ratio
Transmissions AR of a
Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
triangular element of sides l 1 , l 2 , l 3 is defined as:
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose

max (l1 , l2 , l3 )
other than for which it is supplied, without the written consent of AgustaWestland.

AR = (7.11)
min (l1 , l2 , l3 )

The parameters limits specified for the ERICA model superficial mesh are summa-
rized in Table 7.16. The mesh statistics of the finally selected superficial mesh with

Best value Poor at Bad at Worst value


Skewness 0 0.3 0.5 1
Aspect-Ratio 1 1.5 3 1 ∗ 106

Table 7.16: Quality parameters limits specified for ERICA surface mesh.

respect to the above mentioned criteria are summarized in Table 7.17 and Figure
7.52.
7.17 Mesh generation 177

Optimal Poor ele- Bad ele- Worst ele- Mean


elements ments ments ment value
Skewness 175191 3 (0.00 %) 2 (0.00 %) 0.539 0.034
(100 %)
Aspect- 175168 28 (0.02 %) 0 (0.00 %) 2.369 1.144
Ratio (99.98 %)

Table 7.17: Mesh statistic for the finally selected ERICA superficial grid.

As apparent, AgustaWestland quality criteria were fulfilled with wide margins.


Moreover, all the bad elements were located near the tail cone component, in the
region adjacent to the fuselage surface, so that they were expected not to affect the
800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
simulated drag values of the nose, which was the object of the optimization.
Pag. 23 of 98

Figure 8: Histograms representing ERICA surface mesh quality statistics: Aspect-Ratio (on the
Figure 7.52: Histograms representing ERICA(on
left); Skewness surface mesh quality statistics: Aspect-Ratio
the right).
(on the left); Skewness (on the right).

7.2.2 Volume grid

Starting from the surface mesh described above, the volume mesh was generated by means of
7.17.2 Volume grid
the CFD meshing tool Ansys Tgrid®. Actually, the surface mesh created within CATIA® can be
Starting
saved fromcompatible
in a Nastran the surface mesh
format (.dat) described
which can beabove, thebyvolume
easily read Tgrid®. mesh was gener-
atedInby means
Tgrid® of the CFD
environment, meshing
the volume toolthe
between Ansys
tiltrotorTgrid®. Actually,
surface and the surface
the wind-tunnel mesh
walls was
created within CATIA® can be saved in a Nastran compatible format
filled with tetrahedral elements, while the boundary layer region was modelled using prismatic (.dat) which
can be easily read by Tgrid®.
cells.
In Tgrid®
In Table environment,
5, the Tgrid® settings the volume
selected between
to generate the tiltrotor
the internal surface
volume mesh and the wind-
are presented.
tunnel Ten prismatic layers were built over the tiltrotor fuselage in order to ensure that thelayer
walls was filled with tetrahedral elements, while the boundary region
simulated
was modelled
boundary using
layer was prismatic
entirely containedcells.
within In
the Table
prismatic 7.18, the
layers Tgrid®
over settings
the whole selected
fuselage, with the to
generate the internal volume mesh are presented.
exception of regions of separated flow. Moreover, a height of 1 mm was given to the first prismatic
Teninprismatic
layer order to meetlayersthewere built over for
y+ requirements thethetiltrotor fuselage
wall function in order
calculation to ensure
during the CFDthat
the simulated
simulation [4]. boundary layer was entirely contained within the prismatic layers
overAn theimportant
whole aspect
fuselage, with the exception of regions of separated flow. Moreover,
that has to be pointed out is the use of local refinement regions, which
allowed a local improvement of the volume mesh around the ERICA model, in particular near the
nose region, while a coarser tetrahedral mesh could be used in the remaining internal regions. As

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
178 Part II - ERICA nose shape optimization using Ansys Fluent®

Components with boundary layer Canopy, Fuselage, Tail cone


Components without boundary layer Inlet, Outlet, Symmetry
B.L. Offset method Uniform
B.L. Growth method Geometric
B.L. Number of layer 10
B.L. First height 1
B.L. Growth rate 1.2
Tri/Tet Improve surface mesh option Enabled
Tri/Tet Refinment method Adv/front
Tri/Tet Cell size function Geometric
Tri/Tet Growth rate 1.2
Tri/Tet Refinement regions yes

Table 7.18: Ansys Tgrid® settings for volume mesh generation.

a height of 1 mm was given to the first prismatic layer in order to meet the y+
requirements for the wall function calculation during the CFD simulation ([7]). An
important aspect that has to be pointed out is the use of local refinement regions,
which allowed a local improvement of the volume mesh around the ERICA model,
in particular near the nose region, while a coarser tetrahedral mesh could be used
in the remaining internal regions.
As a result, a really good mesh quality was obtained in the regions of interest
(that is, the nose and the cylindrical portion of the fuselage), with the minimum
number of 3D elements. A longitudinal view of the whole volume mesh is de-
picted in Figure 7.53, while Figure 7.54 illustrates a close-up of the mesh near the
tiltrotor surface, where the refinement regions around the fuselage can be clearly
appreciated. Finally, a particular of the boundary layer over the canopy800-91-067
is reported
in Figure 7.55. Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 25 of 98

Figure 9: Longitudinal view of the whole volumetric mesh around the tiltrotor fuselage.
Figure 7.53: Longitudinal view of the whole volumetric mesh around the tiltrotor fuselage.
 

Figure 9: Longitudinal view of the whole volumetric mesh around the tiltrotor fuselage.
7.18 Fluid-dynamic model set up 179

 
800-91-067
Figure 10: Close-up of the volume mesh near the fuselage.
Figure 7.54: Close-up
Numericalofoptimisation
the volumeofmesh
ERICAnear the nose
tiltrotor fuselage.
Rev. A
Pag. 26 of 98

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

Figure 11: Close-up of the volume mesh: boundary layer over the canopy.
Figure 7.55: Close-up of the volume mesh: boundary layer over the canopy.

7.3 Fluid-dynamic model set up


7.18 Fluid-dynamic model set up
A pressure-based solver type with absolute velocity formulation and steady approach was

A adopted
pressure-based solver type
for the tiltrotor simulations. − ω SST
The kwith absolute
model wasvelocity
selected forformulation and
turbulence handling. steady
approach The was
air wasadopted
treated as an for the
ideal gastiltrotor simulations.
having constant specific heats,The − ω SSTenables
whichκautomatically model was
the equation energy resolution. This makes it possible to include the compressibility effects in the
selected for turbulence handling. The air was treated as an ideal gas having
numerical simulations: in fact, as it can be deduced from Table 1, the flight condition selected for
constant specific heats, which automatically enables the equation energy resolution.
optimization features a moderately high Mach number. In Table 6, the air properties assigned for
This makes it possible to include the compressibility effects in the numerical
the ERICA nose optimization are summarized.
simulations: in fact, as it can be deduced from Table 7.14, the flight condition
As far as the solution algorithm is concerned, a SIMPLE scheme was adopted, which solves the
selected for optimization features a moderately high Mach number.
pressure and moment equations separately. A third order MUSCL discretization scheme was
Inselected
Table for 7.19,
all thethe air properties
variables, assigned
since it guarantees a high for the of
accuracy ERICA nosesolution,
the numerical optimization
due to are
summarized.
its potential As far asspatial
to improve the accuracy
solution algorithm
by reducing is concerned,
numerical a SIMPLE
diffusion, particularly for complexscheme
was adopted,
three-dimensionalwhich flows, solves
while not the pressure
negatively andthemoment
affecting equations
total time requested separately. A
for simulations
third when
order MUSCL
compared to thediscretization
second order upwind scheme
scheme. was selected for all the variables, since
it guarantees a high
In particular, for the accuracy
gradient spatialof discretization,
the numerical solution,
the Green-Gauss due
node basedtodiscretization
its potential to
improve spatial accuracy by reducing numerical diffusion, particularly forcell
scheme was chosen instead of the more common Green-Gauss cell based or Least squares complex
three-dimensional
based schemes, because flows,it iswhile not negatively
more suitable for unstructuredaffecting
tetrahedral the
meshtotal
[4], astime requested
is the case

for simulations when compared to the second order upwind scheme. In particular,
for theCopyright
gradient – 2007 spatial discretization,
Agusta SpA, the Ltd,
Westland Helicopters Green-Gauss node based
Westland Transmissions discretization
Ltd and AgustaWestland
scheme was chosen
International instead
Ltd. (Collectively knownof asthe more common Green-Gauss cell based or Least
“AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
180 Part II - ERICA nose shape optimization using Ansys Fluent®

squares cell based schemes, because it is more suitable for unstructured tetrahedral
mesh [7], as is the case for the tiltrotor fuselage simulation.

Gas thermo-dynamic model Ideal Gas


Cp Specific Heat [ J/kg · K ] Constant: 1006.43
Thermal conductivity [W/m · K ] Constant: 0.0242
Viscosity variation law Sutherland: three coefficients methods
µ0 Reference viscosity [kg/m · s] 1.716 · e−5
T0 Reference temperature [K ] 273.11
S Effective temperature [K ] 110.56

Table 7.19: Air properties for ERICA nose optimization.

Solver type Pressure-based steady-state


Pressure-velocity coupling scheme SIMPLE
Spatial discretization schemes
Gradient Green-Gauss node based
Pressure Second order
Density Third-order Upwind
Momentum Third-order Upwind
Turbulent kinetic energy Third-order Upwind
Specific dissipation rate Third-order Upwind
Energy Third-order Upwind

Table 7.20: Solver settings and discretization schemes.

The selected solver settings and discretization schemes are summarized in


Table 7.20.

7.18.1 Boundary and operating conditions


The boundary conditions for the optimization study were selected according to
the operating flight conditions reported in Table 7.14.
A total pressure condition was imposed on the wind tunnel inlet, while a static
pressure was assigned over the outlet section. Over the lateral walls of the wind
tunnel, together with the top and bottom surfaces, a symmetry condition was
imposed. All the other surfaces were treated as hydraulically smooth, adiabatic
walls. The operating reference condition was set to the reference static value for
the prescribed flight condition in Table 7.14, p = 38251.4[ Pa].
The pressure inlet boundary condition imposed on the inlet boundary zone
requires the specification of both the gauge total pressure and the total temperature.
7.18 Fluid-dynamic model set up 181

These quantities were determined applying the ideal gas laws to the data of Table
7.14; in fact, from the static temperature (T∞ in [K], corresponding to the OAT
given in Table 7.14) and the velocity (V∞ , the True air speed) the operating Mach
number (Ma∞ ) was calculated as:

V∞
Ma∞ = √ (7.12)
kRT∞
where k is the specific heats ratio (1.4 for dry air) and R is the gas constant
(287 [ J/(kg · K )] for dry air). The resulting operating Mach number is 0.5806. The
following equations were used to calculate the total pressure and total temperature
at the inlet:
 k −1
k−1

k
2
PT = P∞ 1 + Ma∞ (7.13)
2
k−1
 
2
TT = T∞ 1 + Ma∞ (7.14)
2
being P∞ the static pressure given in Table 7.14. Subtracting the value of the
reference pressure to the total pressure calculated with Equation 7.13, we obtained
a gauge total pressure of 9811.29 [Pa], while the resulting total temperature from
Equation 7.14 was 255.54 [K].
Table 7.21 summarizes the settings for the pressure inlet boundary condition.

Gauge total pressure [ Pa] 9811.29


Supersonic/initial gauge pressure [ Pa] 0
Direction specification method normal to boundary
Turbulence specification method Intensity and length scale
Turbulent intensity [%] 1
Turbulent length scale [m] 0.5
Total temperature [K ] 255.54

Table 7.21: Pressure inlet specification at inlet.

The pressure outlet boundary condition requires the specification of the gauge
static pressure and the back-flow total temperature at the outlet surfaces: a relative
zero gauge pressure (corresponding to the undisturbed value) and a backflow
total temperature equal to the inlet static temperature were chosen. Table 7.22
summarizes the settings for the pressure outlet boundary condition on the outlet
surfaces. Regarding the turbulence specification method, a turbulence intensity
of 1% was selected. As far as the turbulent length scale is concerned, the final
selected value 0.5 [m] was the result of an iterative calculation process. In fact, the
turbulent length scale can be defined as ωκ . Hence, the values of κ and ω were
p

iteratively computed at both inlet and outlet on a CFD model characterized by an


initial turbulent length scale of 1 [m]. However, some tests were performed that
allowed to verify that the tiltrotor nose drag was quite insensitive to variations of
the turbulent length scale, at least for reasonable values of this parameter.
182 Part II - ERICA nose shape optimization using Ansys Fluent®

Gauge pressure [ Pa] 0


Direction specification method normal to boundary
Turbulence specification method Intensity and length scale
Turbulent intensity [%] 1
Turbulent length scale [m] 0.5
Backflow Total temperature [K ] 239.4

Table 7.22: Pressure-outlet specification at outlet.

7.19 Grid sensitivity analysis


In this section, a preliminary sensitivity study which was carried out in order
to identify a suitable mesh to be used during the nose optimization is described.
The final aim was to obtain a reliable CFD model while saving computational time
and resources.
To this purpose, the following strategy was followed:

1. First, the wind tunnel cross-sectional area influence on drag characteristics


was investigated, the superficial mesh and the box length being kept fixed;
the nose drag was computed via a surface integral of both pressure and
viscous drag.

2. Once the proper cross-sectional area was selected, the influence of the box
length was analysed.

3. Next, the superficial mesh impact on aerodynamic performances was investi-


gated and a mesh size was selected, resulting from a trade-off between the
accuracy in predicting aerodynamic drag and the requested CFD calculation
time, which greatly depends on the mesh size.

4. Finally, the influence of the volume mesh refinement on the objective function
value was considered, and the final model was selected.

As far as the wind tunnel cross-sectional area is concerned, a square section


was adopted and several dimensions were investigated, while the box length was
kept fixed to 84 [m]. Moreover, the superficial mesh over the tiltrotor surface was
held unchanged, with an element size of 30 [mm] over the canopy, 40 [mm] over
the cylindrical portion of the fuselage and 80 [mm] on the tail cone.
The analyzed configurations are summarized in Table 7.23, where the nose
drag values are reported for each of the investigated cross-sectional dimensions.
In Figure 7.56, a graphical representation of the nose drag sensitivity to cross-
sectional area is illustrated. As apparent, the nose drag increases with increasing
cross-sectional area, (due mainly to the pressure drag component) until a plateau
is reached. The drag value appears to be stabilized for an 80X80 [m2 ] cross section,
so this model was selected and submitted to further sensitivity analyses.
Wind tunnel Pressure Nose DRAG [N] Viscous Nose DRAG [N] Total Nose DRAG [N]
3
dimensions [m ]
30X30x84 357 666 1024
7.19 Grid sensitivity analysis
40X40x84 526 665 183
1192
60X60x84 627 663 1291
80X80x84
Wind tunnel dimensions [m] 672
Nose pressure DRAG [ N ] Nose 663
vosicous DRAG [ N ] 1336
Total nose DRAG [N]
30x30x84
100X100x84 663 357 663 666 1024 1327
40x40x84 526 665 1192
Table 10:
60x60x84 Drag sensitivity to wind
627 tunnel cross-sectional
663 area dimensions.
1291
80x80x84 672 663 1336
100x100x84 664 663 1327

In Figure 12, a graphical representation of the nose drag sensitivity to cross-sectional area is
Table 7.23: Drag sensitivity to wind tunnel cross-sectional area dimensions.
illustrated.

Figure 12:7.56:
Figure Nose dragdrag
Nose sensitivity to to
sensitivity cross-sectional
cross-sectional area.
area.

As apparent,Next, the nose drag increases


the influence of the with box lengthincreasing was cross-sectional
investigated. Specifically, area, (duestarting mainly to the
pressure drag from a wind-tunnel
component) until length
a plateau equal is to 84 [m], the
reached. Thebox drag length
valuewas increased,
appears to beinstabilized
order for
to2 better reproduce the undisturbed flow at the domain exit, and the effect on the
an 80X80 [m ] cross section, so this model was selected and submitted to further sensitivity
nose drag was monitored. The variations in box length were shown to slightly affect
analyses. nose overall drag values, as apparent from Table 7.24 and Figure 7.57: actually,
Next, the variations
influence of of nose
the box drag are inwas
length the investigated.
order of 10 [N] within the starting
Specifically, considered fromrange of
a wind-tunnel
box lengths. In light of this, it was decided to select the model with the smallest
length equalnumber to 84 [m], of the box length
elements among was theincreased,
four ones in order to better
investigated, which reproduce the undisturbed
is the 80x80x110
3
[m ]. It was characterized by a total number of elements equal to 2.7 millions.
Copyright – 2007 TheAgusta
third step SpA,ofWestland
sensitivity Helicopters
analysis concernedLtd, Westland Transmissions
the superficial meshLtd and AgustaWestland
refinement
Internationallevel. To this purpose,
Ltd. (Collectively knownstarting from the above mentioned element sizes, the mesh
as “AgustaWestland”)
Copyright and all other rights in this documentnose
was refined and the effect on drag was
are vested investigated. Specifically, the new
in AgustaWestland.
superficial
This document contains mesh information
confidential proprietary was characterized
and is supplied on theby ancondition
express element size
that it may ofdisclosed,
not be 20 [mm] over
reproduced theor incanopy,
In whole part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
30 [mm] over the cylindrical portion of the fuselage and 80 [mm] on the tail
cone. The comparison between the two models is summarized in Table 7.25: in
184 Part II - ERICA nose shape optimization using Ansys Fluent®

Wind tunnel dimensions [m] Nose pressure DRAG [ N ] Nose viscous DRAG [ N ] Total nose DRAG [ N ]
BOX4=80x80x110 [m3 ] Length=110 m, 22 m upstream the fuselage 643 663 1316
BOX4=80x80x125 [m3 ] Length=125 m, 40 m upstream the fuselage 636 660 1299
BOX4=80x80x145 [m3 ] Length=145 m, 60 m upstream the fuselage 638 662 1300
BOX4=80x80x180 [m3 ] Length=180 m, 80 m upstream the fuselage 646 660 1306

800-91-067
Table 7.24: Drag sensitivity to wind-tunnel box length.
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 33 of 98

Figure 13: Nose drag sensitivity to wind-tunnel box length.


Figure 7.57: Nose drag sensitivity to wind-tunnel box length.

As apparent, the nose drag change is nearly negligible when refining the superficial mesh over
the first column the superficial grid refinement is reported, with the first number
the nose and cylindrical fuselage. On the other hand, with the refined superficial mesh the total
representing the mesh size of the nose component, the second one the mesh size of
number the
of volume cells
fuselage rises up to and
component, 4.5 millions of elements.
the third Hence,
one the mesh due
size of to computational
the reasons,
tail component. As
the less refined superficial mesh was retained, with 30 [mm], 40 [mm] and 80 [mm] element size
over the nose, cylindrical fuselage and tail cone respectively.
Superficial mesh refinement [m] Nose pressure DRAG [ N ] Nose viscous DRAG [ N ] Total nose DRAG [ N ]
30-40-80 [mm] 643 663 1316
20-30-40 [mm] 644 668 1318
Superficial mesh Pressure DRAG [N] Viscous DRAG [N] Total DRAG [N]
Refinement [mm] Table 7.25: Drag sensitivity to superficial mesh size.
30-40-80 643 663 1316
20-30-80
apparent, the nose drag change 644is nearly negligible 668when refining the 1313
superficial
mesh over the nose Table
and12:cylindrical
Drag sensitivity to superficial
fuselage. On themesh
othersize.
hand, with the refined
superficial mesh the total number of volume cells rises up to 4.5 millions of
elements.
Finally, Hence,
the effects due
of the to computational
volume grid refinement reasons, the less refined
were investigated. It wassuperficial mesh
realized that the
was retained, with 30 [mm], 40 [mm] and 80 [mm] element size over the
volume grid refinement has a great influence on the nose drag. In particular, drag was found to
nose,
cylindrical fuselage and tail cone respectively.
decrease with decreasing growth rate of tetrahedral elements, which governs the transition of the
Finally, the effects of the volume grid refinement were investigated. It was
elements size from the inner region of the domain (near to the prismatic layers) to the external
region. Hence, a sensitivity analysis was performed, in order to highlight the influence of this
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
7.20 Validation of the CFD model 185

realized that the volume grid refinement has a great influence on the nose drag. In
particular, drag was found to decrease with decreasing growth rate of tetrahedral
elements, which governs the transition of the elements size from the inner region of
the domain (near to the prismatic layers) to the external region. Hence, a sensitivity
analysis was performed, in order to highlight the influence of this parameter
on the drag value, keeping the wind tunnel dimensions fixed to 80x80x110 [m3 ]
and holding the superficial grid element size unchanged to the 30-40-80 [mm]
combination. Results are reported in Table 7.26.

Volume mesh refinement Nose pressure DRAG [ N ] Nose viscous DRAG [ N ] Total nose DRAG [ N ]
Growth Ratio: 1.6 643 663 1316
Growth Ratio: 1.2 253 665 918

Table 7.26: Volume mesh refinement influence on the nose drag characteristics.

As apparent, while the viscous drag is nearly insensitive to the volume grid
refinement, the pressure component decreases with increasing refinement. Given
that the number of elements in the latter configuration rises up to 4.4 millions, a
further refinement was not taken in consideration, because a further increase in the
amount of mesh elements would have become prohibitive from the computational
time point of view. Moreover, the validation of the model with growth ratio equal
to 1.2 gave satisfactory results, as will be illustrated in chapter 7.20, so this model
was retained for the optimization study.
Hence, the finally selected model, which was then used in the following vali-
dation analysis, is the 80x80x110 [m3 ] wind tunnel box, 30-40-80 [mm] superficial
element size, with volumetric growth ratio equal to 1.2. Once the model was
selected, the y+ characteristics of the associated CFD solution were investigated.
In Figure 7.58, the y+ distribution over the tiltrotor fuselage is illustrated.
As apparent, the selected parameters for grid generation were shown to guar-
antee that the non-dimensional mesh thickness at the fuselage surface fell within
the discretization levels (y+ = 30 ÷ 300) suggested for the standard wall functions
implemented in the conventional turbulence models to work properly, in partic-
ular over the nose region, where the fluid-dynamic quantities necessary for the
optimization objective function evaluation are calculated.

7.20 Validation of the CFD model


Some experimental data on a non-powered 1/8 scaled ERICA tilt-rotor modular
model were available from a series of low speed wind tunnel tests performed
in the framework of WP4.4 of the NICETRIP Project (FP6/Aeronautics project
AIP5-CT-2006-030944).
The experimental campaign was held at Politecnico di Milano and split into
three major phases, in which different model configurations were tested. An
exhaustive analysis of the acquired aerodynamic coefficients of both the whole
800-91-067
186 Part II Numerical
- ERICA nose shape optimization
optimisation using
of ERICA tiltrotor Ansys
nose Fluent®
Rev. A
Pag. 35 of 98

FigureFigure Contours
7.58:14: Contoursof
ofwall y+over
wall y+ overthethe ERICA
ERICA fuselage.
fuselage.

aircraft and the isolated components was carried out in [73]. Specifically, the
contribution of each component to the model global aerodynamic coefficients was
9. Validation of the CFD model
highlighted, with the main objective of investigating the aircraft drag breakdown.
Actually, the above data
Some experimental mentioned experimental
on a non-powered observations
1/8 scaled refer modular
ERICA tilt-rotor to specified
modelvalues
were
of Reynolds
available from aand Mach
series of lownumbers
speed windtypical
tunnel of theperformed
tests wind tunnel environment
in the framework and ofofthe
of WP4.4 the
scaled model
NICETRIP dimensions.
Project In [73],
(FP6/Aeronautics anAIP5-CT-2006-030944).
project extrapolation of theThe wind tunnel measured
experimental campaign
dataheld
was on the model scaled
at Politecnico tilt-rotor
di Milano and splitwas
into carried outphases,
three major in order to infer
in which somemodel
different basic
aerodynamic characteristics of the full scale aircraft: viscosity and compressibility
configurations were tested. An exhaustive analysis of the acquired aerodynamic coefficients of
effects
both the were
whole accounted for isolated
aircraft and the by means of some
components wasempirical
carried outcorrelations published
in [10]. Specifically, the
in the literature. Starting from the experimental measurements, the lift
contribution of each component to the model global aerodynamic coefficients was highlighted, with
and drag
coefficients of the main aircraft components in airplane mode were derived at
the main objective of investigating the aircraft’s drag breakdown. Actually, the above mentioned
typical full scale cruise Reynolds and Mach numbers. Then, their cumulative effect
experimental observations refer to specified values of Reynolds and Mach numbers typical of the
on the global aerodynamic coefficients of the aircraft was evaluated.
wind tunnel environment and of the scaled model dimensions. In [11], an extrapolation of the wind
In this chapter, a validation of the tiltrotor CFD model selected for the opti-
tunnel measured data on the model scaled tilt-rotor was carried out in order to infer some basic
mization runs was performed against wind tunnel data over the bare fuselage.
aerodynamic characteristics of the full scale aircraft: viscosity and compressibility effects were
Specifically, the geometry and mesh were scaled down to the actual 1/8 wind
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
tunnel model dimensions and the fluid dynamic simulation was carried out at
International Ltd. (Collectively known as “AgustaWestland”)
the experimental conditions, with an incidence angle equal to that selected for
Copyright and all other rights in this document are vested in AgustaWestland.
theThisoptimization
document contains confidentialruns.
proprietary Then,
information andthe simulated
is supplied fuselage
on the express condition that it may not drag
be disclosed,was compared
reproduced In whole or in part, or with
used for anythe
purpose
other than for which it is supplied, without the written consent of AgustaWestland.
experimental acquisitions. As will be illustrated in the following, a good agreement
of the calculated drag with experiments was found at wind tunnel conditions.
When taking into account the effects of both full-scale Reynolds and Mach number
at real flight conditions on the aerodynamic coefficients, this simulation served as
a validation of the numerical model to be used in the optimization phase. Actually,
being the fuselage drag at wind tunnel conditions captured with a satisfactory
7.20 Validation of the CFD model 187

accuracy, the CFD model was judged capable of properly reproducing the tiltrotor
aerodynamic behaviour even at the full-scale conditions proper of the optimization
runs.
In Table 7.27, the flowfield characteristics of the selected test case for validation
at model-scaled conditions are summarized, in terms of flow static pressure and
temperature, Reynolds number referred to the wing mean aerodynamic chord and
Mach number.

Static pressure [ Pa] 97956


Static temperature [K ] 294.18
Air density [kg/m3 ] 1.16
Air speed [m/s] 44.85
Rewingchord 8.73751E+05
Mach 0.13

Table 7.27: Flowfield characteristics at wind tunnel conditions.

The same fluid dynamic set up of the optimization runs was used for the
model validation. In particular, a pressure-based solver type with absolute velocity
formulation and steady approach was selected for the simulations. The κ − ω SST
model was chosen as the viscous model with turbulence intensity equal to 1%, and
the air was treated as an ideal gas having constant specific heats. The boundary
conditions were as follows: a total pressure condition was imposed on the wind
tunnel inlet, while a static pressure was assigned over the outlet section, the actual
values of both depending on the data of Table 7.27: to all the lateral surfaces of the
wind tunnel and to its top and bottom surfaces a symmetry condition was applied.
The fuselage surface was treated as a hydraulically smooth, adiabatic wall.
The adopted boundary conditions for the model-scaled case are summarized
in Table 7.28.

Model Scale
total pressure [ Pa] 99119.72
INLET static pressure [ Pa] 97956
total temperature [K ] 295.2
static pressure [ Pa] 97956
OUTLET
backflow total temperature [K ] 295.2

Table 7.28: Boundary conditions on the fluid domain for the validation case.

As far as the solution algorithm is concerned, a SIMPLE scheme was adopted,


along with a third order MUSCL discretization scheme for all the variables. A
normalized residual on the continuity equation around 1 ∗ 10−6 was obtained at the
end of the simulation run and both the fuselage lift and drag coefficients reached
stabilized values.
188 Part II - ERICA nose shape optimization using Ansys Fluent®

In Table 7.29, the fuselage drag coefficients of the model-scaled tiltrotor coming
from the simulations is reported and compared to the experimental one. The
drag was predicted with a satisfactory accuracy, leading to a percentage error
of 9 % with respect to the wind tunnel value, which is considered good for the
optimization purposes.
In Figure 7.59, the experimental drag curve of the tiltrotor fuselage is depicted,
along with the obtained CFD results.

Experiment at α = −1.99[deg] Simulation at α = −1.97[deg] % deviation


Model-scaled 0.01883 0.017 -9[%]
Drag coefficient
Reference Area= 0.578 [m2 ]

Table 7.29: Model-scaled tiltrotor fuselage drag coefficient: simulations vs. experiments.
800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 38 of 98

CD of the tiltrotor fuselage
0.070

0.060

0.050
Experiment
0.040 Simulation

CD
0.030

0.020

0.010

0.000
‐20 ‐10 0 10 20 30
alpha [deg]  

Figure 15: Model-scaled tiltrotor fuselage drag curve: simulations vs. experiments.
Figure 7.59: Model-scaled tiltrotor fuselage drag curve: simulations vs. experiments.

Given that the correlation with model-scaled experiment was accurately cap-
tured, the CFD model was judged accurate enough also when transposing the
results at the higher Reynolds and Mach number proper of the operating flight
conditions, at which the nose optimization was carried out.
7.21 Parameterization 189

7.21 Parameterization
Once the baseline CFD solution was available, the geometry parameterization
was carried out. This operation is of outstanding importance in the optimiza-
tion process and it can be performed using various techniques. For the scope
of the present work, the commercial software HyperMesh® was chosen as the
parameterization tool, due to the really strong and versatile capabilities of the
morphing tool HyperMorph®. The choice of this software is mainly justified with
its effectiveness and ease of use, which allows the user to build up complicated
parametric models in a short time by means of a graphical user interface. Moreover,
the parameterization procedure in HyperMorph® is very general, in the sense that
it is independent from the peculiar model (either FEM or CFD) which is the object
of the optimization analysis. Therefore, whatever the geometry, the user is always
allowed to use the same morphing strategies, again saving a lot of working time.
Both those characteristics are fundamental in an industrial context, where time is
always an essential issue.
The strategy chosen to correctly perform the parameterization can be summa-
rized with the following steps:

1. Importation of the baseline case file (i.e., the model 30-40-80, 80x80x110,
growth ratio 1.2) into HyperMesh® in Nastran Format.

2. Creation of the desired shapes, starting from the baseline geometry, through
the morphing techniques available within HyperMorph®. For CFD studies,
the more suitable strategy is the domains-handles approach to perform
localized deformation, and the morph-volumes approach to satisfy global
morphing requirements.

3. Saving of the generated shapes; with this operation, HyperMorph® stores the
current handles − nodes perturbations allowing the user to apply them to
the undeformed model with any given scaling factor. With this method, the
scaling factor of any generated shape can be dealt with as a design variable
by an optimization algorithm.

4. Declaration of the created shapes as design variables for the HyperStudy


optimization analysis, following the guidelines presented in [4].

5. Saving of the current parameterized model into an .hm file, which will be
then exploited for the batch-mode parameterization.

6. Creation of the HyperMesh® command file. An example of such a command


file can be found in Appendix C.3; its function will be described in more
detail in chapter 7.22.

7.21.1 Morphing Constraints


The only component to be optimized in the present study was the tiltrotor
nose: hence, both the cylindrical portion of the fuselage and the tail cone were kept
6) Creation of the HyperMesh® command file, that can be found in Appendix A.1; its
function will be described in more detail in chapter 11.

190 Part II Constraints


10.1 Morphing - ERICA nose shape optimization using Ansys Fluent®
The only component to be optimized in the present study was the tiltrotor nose: hence, both the

unchanged. In other
cylindrical portionwords, all the
of the fuselage and nodes over
the tail cone both
were kept the fuselage
unchanged. and
In other thealltail
words, the cone

surface (purple dots in Figure 7.60) were constrained to stay fixed.


nodes over both the fuselage and the tail cone surface (purple dots in Figure 16) were constrained
to stay fixed.
Even though it slightly increased the computational resources required during
Even though it slightly increased the computational resources required during the mesh
the mesh deformation
deformation operations, the introduction of the morphing constraints was
operations, the introduction of the morphing constraints was mandatory to obtain a
mandatory to obtain a suitable optimized geometry, conformal to non-aerodynamic
suitable optimized geometry, conformal to non-aerodynamic requirements, such as manufacturing
requirements, suchconstraints.
and structural as manufacturing and structural constraints.

Figure 16: Fixed nodes over the fuselage and tail surfaces.
Figure 7.60: Fixed nodes over the fuselage and tail surfaces.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland

In addition, tangency
International constraints
Ltd. (Collectively between the nose component and the cylin-
known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
drical portion of the fuselage were introduced, by means of the Set Biasing option
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

made available within Hypermorph® environment.

7.21.2 Design Variables Definition


In the following, a brief description of each design parameter taken into
consideration for the nose optimization study is provided.
Obviously, some parameters are symmetrical with respect to the ERICA main
axis; thereby they will be presented together, since they were deformed simultane-
ously during both D.O.E. analysis and the optimization study.
A total amount of nine design variables were considered. However, after the
Student analysis only eight of them were shown to be influential on the nose drag,
and they were retained for the optimization procedure.
All the design variables, with the exception of the sh4, were built by means
of the morph-volumes approach in order to satisfy global morphing requirements.
Moreover, with this approach more streamlined shapes of the nose could be
achieved than those obtainable with the domains-handles approach. The following
design variables were identified for the nose optimization:

1. sh1-sh2: this was a symmetrical shape, which consisted of a rotation of the


first eight handles of the morph volume block around two symmetrical
axes, as illustrated in Figure 7.61; in this figure, the baseline shape along
with the morph volumes created around the nose are presented. The initial
deformation range assigned to this variable was equal to ±30 [deg];
7.22 D.O.E. and Student analysis 191

2. sh3: this shape consisted of the nose stretching along the negative direction
of the x-axis. The deformation range was given a value of 0 ÷ 0.5[m]. It is
shown in Figure 7.62;

3. sh4: this shape consisted of an initial deformation along z-axis in the nose
region near the windscreen (Figure 7.63). It was given an initial deformation
range of ±0.15 [m], due to pilot visibility constraints;

4. sh5: this shape consisted of the rotation of two morph-volume handles,


located at the interface between the nose and the cylindrical portion of the
fuselage, as illustrated in Figure 7.64. It was given an initial deformation
range of ±5 [deg], due to pilot visibility reasons.

5. sh6-sh7: this was a symmetrical shape, consisting of a rotation of two morph-


volume handles along two symmetrical axes, located at the interface between
the nose and fuselage (Figure 7.65). It was given an initial deformation range
of ±20 [deg], due to pilot visibility reasons;

6. sh8-sh9: this shape was similar to the sh1-sh2 one, but a wider portion of
the canopy was involved in this case: in fact, the rotational axes were moved
upstream (Figure 7.66).The initial deformation range assigned to this variable
was equal to ±15 [deg];

7. sh10-sh11: this was a symmetrical shape, conceived to deform the nose region
along y-axis, as shown in Figure 7.67. It was given an initial deformation
range of ±0.3[m];

8. sh12-sh13: this shape was similar to the previous one, but it regarded a
portion of the nose located downstream. The initially selected deformation
range was equal to ±0.1 [m]. It is shown in Figure 7.68 (the arrows are small
when compared to the ones of the previous shape, since they are proportional
to the shape displacement);

9. sh14-sh15: this was a shape located at an intermediate position between the


previous ones, that is sh10-sh11 and sh12-sh13, as apparent in Figure 7.69. It
was given an initial deformation range of ±0.2 [m].

7.22 D.O.E. and Student analysis


Once both the sensitivity analysis and the geometry parameterization were
completed, the Design of Experiment study loop was performed.
Design of Experiments (D.O.E.) can be defined as a test or a series of tests in
which the input variables of a process or system are changed so that the reasons
for changes in the output response can be identified and observed.
In this work, D.O.E. was applied with the aim of exploring the design space
and determining which factors were most influential on the response. To this
purpose, a Student analysis was performed exploiting the results of D.O.E.
axis, as shown in Figure 23. It was given an initial deformation range of ± 0.3[m];
8) sh12-sh13: this shape was similar to the previous one, but it regarded a portion of the
nose located downstream. The initially selected deformation range was equal to ± 0.1
[m]. It is shown in Figure 24 (the arrows are small when compared to the ones of the
192 previous Part
shape,II - ERICA
since nose shape
they are proportional to theoptimization using
shape displacement); Ansys Fluent®
9) sh14-sh15: this was a shape located at an intermediate position between the previous
ones, that is sh10-sh11 and sh12-sh13, as apparent in Figure 25. It was given an initial
deformation range of ± 0.2 [m].

 

Figure 17: Shape sh1-sh2 definition.


Figure 7.61: Shape sh1-sh2 definition.

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
800-91-067
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 43 of 98

Figure 18: Shape sh3 definition.


Figure 7.62: Shape sh3 definition.

7.22 D.O.E. and Student analysis 193
Figure 18: Shape sh3 definition.

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A 
Pag. 44 of 98
Figure 19: Shape sh4 definition.
Figure 7.63: Shape sh4 definition. 800-91-067
Copyright – 2007 Agusta SpA, Westland
NumericalHelicopters
optimisationLtd, Westland
of ERICA Transmissions
tiltrotor nose Ltd
Rev. A and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”) Pag. 44 of 98

Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

Figure 20: Shape sh5 definition.


Figure 7.64: Shape sh5 definition. 

Figure 20: Shape sh5 definition.



Figure 21: Shape sh6-sh7 definition.


Figure 21: Shape sh6-sh7 definition.
Figure 7.65: Shape sh6-sh7 definition.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
CopyrightInternational
– 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
Ltd. (Collectively known as “AgustaWestland”)
International Ltd. (Collectively
Copyright and all other known
rights inas
this“AgustaWestland”)
document are vested in AgustaWestland.
Copyright This
and document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
all other rights in this document are
other than for which it is vested in the
supplied, without AgustaWestland.
written consent of AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
194 Part II - ERICA nose shape optimization using Ansys Fluent®

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 45 of 98

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 45 of 98

Figure 22: Shape sh8-sh9 definition.


Figure 7.66: Shape sh8-sh9 definition.

Figure 22: Shape sh8-sh9 definition.

Figure 23: Shape sh10-sh11 definition.

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland. 

Figure 23: Shape sh10-sh11 definition.


Figure 7.67: Shape sh10-sh11 definition.

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
800-91-067 800-91-067
Numerical
Numerical optimisation
optimisation of ERICAoftiltrotor
ERICAnose
tiltrotor Rev.
nose A Rev. A
7.22 D.O.E. and Student analysis 195
Pag. 46 of 98Pag. 46 of 98



Figure Figure 7.68:
24: Shape Shape definition.
sh12-sh13 sh12-sh13 definition.
Figure 24: Shape sh12-sh13 definition.

Figure 25: Shape sh14-sh15 definition. 

Figure 25: Shape sh14-sh15 definition.


Copyright – 2007 Agusta SpA, Westland Helicopters
Figure 7.69: Ltd,
ShapeWestland Transmissions
sh14-sh15 Ltd and AgustaWestland
definition.
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other
Copyright rights
– 2007 in this SpA,
Agusta document are vested
Westland in AgustaWestland.
Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
TheLtd.
International basic operations
(Collectively knownperformed within the
as “AgustaWestland”) D.O.E. loop are illustrated in the
other than for which it is supplied, without the written consent of AgustaWestland.

central
Copyright and part of the
all other rightsflow chart
in this reported
document in Figure
are vested 7.45. Each program involved in the
in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
loop is interfaced and other synchronized
than for which it is supplied, without the written consent ofso
with the others that, when one program is
AgustaWestland.

running, the following one waits until the preceding execution is complete. This
allows to generate an input/output chain, in which each program uses the output
of the preceding software as its own input.
196 Part II - ERICA nose shape optimization using Ansys Fluent®

The input/output chain illustrated in Figure 7.45 works as follows:

1. The D.O.E. Matlab® script, compiled in Matlab® environment using a ded-


icated command, creates the actual values of the design variables to be
processed. A Latin Hypercube D.O.E. was selected in this work, which sam-
ples the entire design space into equal-probability regions, while allowing the
designer to decide the amount of samples to be performed. A total number
of 63 samples was selected, in order to keep the overall computational time
to a reasonable level.

2. The design variables set are then imported, one at a time, into HyperMesh®,
which, in turn, performs the shape parameterization according to the received
values by means of its dedicated tool HyperMorph®. To this purpose, a tcl
file Command.tcl was prepared to carry out these operations. The command
that allows to launch HyperMesh® in batch mode in Windows® environment
is the following:

\emph{[path]/hmbatch.exe -c<Command.tcl>

where hmbatch.exe is the batch-mode release of HyperMesh® (executable on


Windows® platform), which performs shape parameterization in batch-mode,
whereas Command.tcl is the name of the command file. HyperMesh® opens a
previously created <filename.hm> file, containing the superficial mesh, along
with the design variables built using HyperMorph® (please see chapter 7.21);
then, the superficial deformations are carried out, and a Nastran format file
(<filename.dat>) containing the deformed superficial mesh is created.

3. <filename.dat> is then transferred to Ansys Tgrid® software, which creates


the volume mesh according to the instructions contained in a Ansys Tgrid®
journal file, Tgrid-batch.jou, which can be found in Appendix C.4. As a re-
sult, a <filename.cas> is built, which is then submitted to Ansys Fluent®
calculations;

4. Ansys Fluent® performs the CFD run and returns the objective function
values to the D.O.E. code, which, in turn, provides Altair HyperMesh® with
the next set of design variables; then, the loop starts again. A dedicated
journal file, Fluent-batch.jou, is used to launch Ansys Fluent® in batch mode:
the journal file can be found in Appendix C.5.

In the present work, each variable was given a normalized deformation range
equal to ±1, where +1 corresponds to the upper variable value and -1 to the lower
variable value (the true range of each variable was indicated in chapter 7.21).
Once all the calculations have been performed, the sets of the design variables,
along with the correspondent values of the nose drag (computed in Ansys Fluent®
environment), were transferred to the Matlab® Student.m file. Student.m file, which
supervises the entire Student analysis block, can be found in Appendix C.7.
7.22 D.O.E. and Student analysis 197

Actually, the most influential factors on the objective function were determined
using a statistical analysis based on the t-Student parameter:

| x1 − x2 |
t= (7.15)
σ
being x1 the mean value of the objective function for the upper set, namely
the set of configurations where the investigated variable is given the upper value
(i.e., from the mean value, 0, to the upper value, +1); x2 is the mean value of
the objective function for the lower set, (i.e., the set of configurations where the
investigated variable is given the lower values, from 0 to -1), and σ the standard
deviation, defined as follows:
v
u ∑n1 ( x1i − x¯1 )2 + ∑n2 ( x2i − x¯2 )2 (n1 + n2 )
u 
i =1 i =1
σ= (7.16)
t
( n1 + n2 − 2) n1 n2

where x1i and x2i are the objective functions of the ith configuration of the upper
and lower set respectively, and n1 and n2 are the cardinalities of the upper and
lower configuration sets.
As it is well known, the t-Student parameter assesses whether the means of two
groups are statistically different from each other, and it is defined in such a way
that the bigger its value, the higher the difference between the two populations, and
hence the higher the response variation caused by the corresponding parameter.
Therefore, a design variable with an associated high value of t-Student parameter
is expected to have a higher influence on the response than a variable with a lower
800-91-067
Student parameter. The results of optimisation
Numerical the Studentofanalysis on the
ERICA tiltrotor tiltrotor
nose nose drag
Rev. A
are reported in Figure 7.70. Pag. 50 of 98

Figure 7.70: Student


Figure analysis
26: Student ononthe
analysis thedesign variables
design variables of ERICA
of ERICA nose
nose total total drag.
drag.

Design Variable Baseline's values Upper range Lower range


sh1-sh2 0 -0.2 1

sh3 0 0 0.5

sh4 0 -0.2 0.5

sh5 0 0 0.2

sh6-sh7 0 0 0

sh8-sh9 0 -0.2 0.5


198 Part II - ERICA nose shape optimization using Ansys Fluent®

As apparent, the fifth design variable, corresponding to shapes sh6-sh7 (please


see chapter 7.21) is expected to have a small influence on the ERICA nose total
drag. Thereby, it was decided to discard it in the following optimization study.
Moreover, since the Student analysis showed that the initial selected deformation
ranges could yield non-streamlined shapes, it was decided to modify the variation
range of each variable. Specifically, the variables were given a non-symmetric range
and the allowed deformations were reduced with respect to the initial ones. Finally,
different deformation ranges were assigned to each variable, based on the Student
analysis results.
The newly defined deformation ranges are summarized in Table 7.30.

Design variable Baseline values Lower range Upper range


sh1-sh2 0 -0.2 1
sh3 0 0 0.5
sh4 0 -0.2 0.5
sh5 0 0 0.2
sh6-sh7 0 0 0
sh8-sh9 0 -0.2 0.5
sh10-sh11 0 -0.2 0.5
sh12-sh13 0 -0.2 0.5
sh14-sh15 0 -0.2 0.8

Table 7.30: Design variables deformation ranges.

7.23 Optimization study: discussion of results


Once the design variables to be retained and the proper deformation ranges
were identified through both the D.O.E. analysis and the Student test, the opti-
mization study was performed, using the GeDEAII algorithm.
The basic operations proper of this step of the optimization procedure are
illustrated in the lower part of the flow chart reported in Figure 7.45, that is
the GeDEAII driven optimization block. Actually, they are very similar to the
operations performed during the D.O.E. analysis, with the exception of the block
master code (highlighted in yellow in Figure 7.45). In fact, the supervising code in
this phase is the GeDEAII algorithm.
The Matlab® function which performs all the synchronization instructions is
called Command-function.m, and can be found in Appendix C.2. The remainder of
the GeDEAII code was left unchanged.
The involved softwares are Ansys Fluent® for the CFD calculations and the
objective function computation, Ansys Tgrid® for the mesh generation, and Altair
Hypermesh® as the parameterization tool. The total number of individuals per
each generation was set to 12, and the initial randomly generated population
evolved for 3 generations, which were considered sufficient for a single-objective
optimization. The optimization convergence history is depicted in Figure 7.71.
7.23 Optimization study: discussion of results 199

The evolution process resulted in an optimized configuration characterized by

920
Total drag

910

900

Total drag [N]


890

880

870

860
0 1 2 3
Number of generations

Figure 7.71: Convergence history of the GeDEAII driven optimization.

a drag reduction of nearly 6%, when compared to the baseline configuration.


A comparison between the baseline and the optimized geometries is depicted
in Figure 7.72, where a sketch of both the configurations on both a top and a
longitudinal view is represented.

The objective function values for both the baseline and the optimized nose
geometries are reported in Table 7.31, where the total drag, along with the pressure
and viscous components are reported. As apparent, the optimization has led to a
significant reduction of the pressure drag, while keeping the viscous component
substantially unchanged. The design variables values of the optimized configura-
tion are summarized in Table 7.32, where both the absolute and normalized values
are reported (one has to remember that a normalized value equal to 1 means that
the corresponding variable assumes the upper allowed value of the range). Except
for design variables sh3 and sh8-sh9, all the other variables were given a value near
200 Part II - ERICA nose shape optimization using Ansys Fluent®

Figure 7.72: Geometrical comparison between the baseline and the optimized nose config-
urations: longitudinal view (top) and top view (bottom).

Total nose drag [N] Pressure component [N] Viscous component [N]
Baseline 918 253 665
Optimized 862 192 669

Table 7.31: Objective function values for the baseline and optimized nose configurations.

to the maximum allowed in the optimized configuration. In particular, sh4, sh5,


sh10-sh11, sh12-sh13 and sh14-sh15, were equal to the upper value in their range.
This is due to the fact that the above mentioned variables had a positive effect
on the reduction of pressure drag, since they acted in the sense of reducing the
cross sections of the nose component. It is also worth considering that parameter

sh1-sh2 sh3 sh4 sh5 sh6-sh7 sh8-sh9 sh10-sh11 sh12-sh13 sh14-sh15


Absolute 0.9077 0.1172 0.5 0.2 0 0.255 0.5 0.4876 0.8
Normalized 0.9231 0.2344 1 1 0 0.65 1 0.98 1

Table 7.32: Absolute and normalized design variables values of the optimized configura-
tion.

sh3, which governs the nose deformation along the longitudinal x axis, takes an
intermediate value between lower and upper limits. This can be explained as long
as a compromise between pressure and skin friction drag components is to be
achieved when the minimum drag is searched for. The pressure coefficient and skin
friction coefficient contours over both the baseline and optimized nose geometries
are illustrated in Figure 7.73 and Figure 7.74 respectively. A series of cuts over
both the baseline and the optimized geometry were performed along the planes
depicted in Figure 7.75 in order to better analyze the pressure and viscous drag
7.23 Optimization study: discussion of results 201

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 55 of 98

pressure drag between x=1.5 m and x=3 m. However this does not affect the overall pressure
drag, which is in fact reduced.

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 55 of 98

pressure drag between x=1.5 m and x=3 m. However this does not affect the overall pressure
drag, which is in fact reduced.

Figure 29: Pressure coefficient contours over the baseline configuration (left) and the optimized
Figure 7.73: Pressure coefficient contours over
configuration the baseline configuration (left) and the
(right).
optimized  configuration (right).

Figure 29: Pressure coefficient contours over the baseline configuration (left) and the optimized
configuration (right).
 

Figure 30: Skin friction coefficient contours over baseline configuration (left) and the optimized
configuration (right).

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

Figure 30: Skin friction coefficient contours over baseline configuration (left) and the optimized
Figure 7.74: Skin friction coefficient contours(right).
configuration over baseline configuration (left) and the
optimized configuration (right).
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
202 Part II - ERICA nose shape optimization using Ansys Fluent®

components on the fuselage nose and identify the reasons of improvement of the
objective function featured by the optimized configuration. To this purpose, the
local values of both pressure and viscous drag of the optimal solution over the
above mentioned planes are illustrated in Figure 32 and Figure 33 respectively,
and compared with the baseline geometry. These values were obtained using the
following formulae:

D p ( x ) = pDA x (local pressure drag) (7.17)

(7.18)

Dv ( x ) = τx DAy + DAz (local skin − f riction drag)
where p is the local value of pressure, DA x the local surface Area projection
normal to the x-axis, τ is the local shear stress (resolved along its x-axis component),
and DAy and DAz the local surface Area projection normal to the y-axis and z-
axis, respectively. As apparent, the zones where pressure and viscous drag are
generated can be deduced by analyzing the behaviour of D p and Dv . From Figure

 
PLANE F 
PLANE E
 
30 deg

PLANE D

PLANE C
5 deg 
5 deg  PLANE B

30 deg 

PLANE A

Figure 7.75: Visualization of the planes selected for pressure and viscous drag analysis.

7.76, it can be argued that the main difference between original and optimized
configuration, in terms of pressure drag, originates close to planes “C” and “D”,
where smaller values in the pressure drag can be found between x=1.5 m and x=3
m on the pressure side of the nose. On the contrary, in plane “F” the optimized
solution exhibits slightly higher values of the pressure drag between x=1.5 m and
x=3 m. However this does not affect the overall pressure drag, which is in fact
reduced. On the other hand, from Figure 7.77, it can be deduced that the original
and optimized configuration are almost equivalent in terms of viscous drag. In fact,
the curves over the planes “A” and “E” are almost coincident, while on plane “F”
the optimized solution exhibits higher values of the viscous drag between x=1.5 m
800-91-067
Numerical
7.23 Optimization study: optimisation
discussion of ERICA tiltrotor nose
of results Rev. A
203
Pag. 57 of 98

PLANE A PLANE B
25 25

Pressure drag ‐x component [N]
Pressure cdrag‐x component [N]
20 20
OPTIMIZED
OPTIMIZED BASELINE
15 15
BASELINE

10 10

5 5

0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12

‐5 ‐5
x [m] x [m]

PLANE C PLANE D
25 25
Pressure cdrag‐x component [N]

Pressure drag ‐x component [N]
20 20

OPTIMIZED
OPTIMIZED
15 BASELINE 15
BASELINE

10 10

5 5

0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12

‐5 ‐5
x [m] x [m]

PLANE E PLANE F
25 25
Pressure cdrag‐x component [N]
Pressure cdrag‐x component [N]

20 20

OPTIMIZED OPTIMIZED
15 15 NASELINE
BASELINE

10 10

5 5

0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12

‐5 ‐5
x [m] x [m]

Figure 32: Pressure drag component over the selected planes: comparison of baseline and
Figure 7.76: Pressure drag component
optimizedover the selected planes: comparison of baseline
configurations.
and optimized configurations.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

and x=3 m. However, over planes “B”, “C” and “D” the viscous component of the
optimized solution is smaller than the baseline between x=1.5 m and x=4 m and
slightly higher between x=4m and x=5.5 m. This results in an overall compensation
of the viscous drag, which is in fact nearly unchanged from the baseline to the
optimized solution.
204 Part II - ERICA nose shape optimization using Ansys Fluent®

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 57 of 98

PLANE A PLANE B
25 25

Pressure drag ‐x component [N]
Pressure cdrag‐x component [N]

20 20
OPTIMIZED
OPTIMIZED BASELINE
15 15
BASELINE

10 10

5 5

0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12

‐5 ‐5
x [m] x [m]

PLANE C PLANE D
25 25
Pressure cdrag‐x component [N]

Pressure drag ‐x component [N]

20 20

OPTIMIZED
OPTIMIZED
15 BASELINE 15
BASELINE

10 10

5 5

0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12

‐5 ‐5
x [m] x [m]

PLANE E PLANE F
25 25
Pressure cdrag‐x component [N]
Pressure cdrag‐x component [N]

20 20

OPTIMIZED OPTIMIZED
15 15 NASELINE
BASELINE

10 10

5 5

0 0
0 2 4 6 8 10 12 0 2 4 6 8 10 12

‐5 ‐5
x [m] x [m]

Figure 32: Pressure drag component over the selected planes: comparison of baseline and
Figure 7.77: Viscous drag component overconfigurations.
optimized the selected planes: comparison of baseline and
optimized configurations.
Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
7.24 Reverse Engineering Procedure in CATIA® V5 205

7.24 Reverse Engineering Procedure in CATIA® V5


From the optimization run the values of the design variables that minimize the
nose drag were derived. Those values were introduced within Altair Hypermesh®
in order to obtain the morphed mesh of the optimal geometry. Then, the mesh was
exported in stl format: specifically, only the modified components were exported,
being the cylindrical portion of the fuselage and the tail cone kept unchanged.
Moreover, since the geometry is symmetrical with respect to the zx plane,
only half of the model was reconstructed. The reverse engineering process is
qualitatively described in the following.
The file *.stl created in Altair Hypermesh® was read within CATIA® V5 by
using the Import function under the Digitalize Shape Editor module. The reader
is referred to Figure 7.49 for the orientation of the coordinate axes used in the
following.
Several planes normal to the x-axis needed to be created in order to extract
the pertinent geometry sections, their number depending on the complexity of
the original model. In this case, four planes were generated, whose coordinates in
the x direction are summarized in Table 7.33. In particular, plane 5 represents the

Plane 1 Plane 2 Plane 3 Plane 4 Plane 5 Plane 6 Plane 7 Plane 8 Plane 9


Absolute origin 1420 1700 2100 2700 3165 3500 4265 5065 5500

Table 7.33: x-coordinate of the generated planes normal to x-axis.

discontinuity interface between the nose and the windscreen. In addition to planes
normal to the x direction, three more planes were created:

• Plane 10: it derives from a rotation of -1.97 deg (equal to the fuselage inci-
dence) of the xy plane around the y-axis and passes through the point at the
top of the nose;

• Plane 11: it is a clockwise rotation of 20° of plane 10 around the x-axis;

• Plane 12: it is a clockwise rotation of 45° of plane 10 around the x-axis;

First, the Planar Section tool was used in order to extract sections of the stl
model over the above mentioned planes. Then, using the Curve from Scan tool,
the pertinent curves were generated with a defined tolerance and with the least
possible number of segments of the least possible order. The tolerance value must
be accurately selected, since a too low value may detrimentally affect the quality
of the generated curves and consequently of the derived surfaces. During curve
creation, the user may select particularly meaningful points, for instance the point
at the top of the nose or the intersection between plane 5 and the zx plane. In the
following, the curves generated from sections 1 to 9 will be referred to as group 1,
while those generated from planes 10 to 12 will be referred to as group 2.
curves generated from sections 1 to 9 will be referred to as group 1, while those generated from
planes 10 to 12 will be referred to as group 2.
The group 1 curves were then split using planes zx, 10, 11 and 12: this operation is represented
in Figure 34.
206 Part II - ERICA nose shape optimization using Ansys Fluent®

Figure 34: The reverse engineering procedure: cyan curves represent the group 1, while black
Figure 7.78: The reverse engineering procedure: cyan curves represent the group 1, while
curves represent the group 2.
black curves represent the group 2.

In addition to these operations, the region between the nose and the windscreen was handled in
The group 1 curves were then split using planes zx, 10, 11 and 12: this operation
a peculiar way. As a matter of fact, in order to respect the original variation of curvature between
is represented in Figure 7.78.
the nose and the windscreen surfaces, some points were added to these curves. A particular of
In addition to these operations, the region between the nose and the windscreen
this operation regarding
was handled in curve generated
a peculiar way. using plane 11 is depicted in Figure 35. A spline was
created using the Asabove
a matter mentioned
of fact, points,
in order thatto isrespect
representedthe original in Figure 36 (colored
variation in red), along
of curvature
with the other between
construction the nose lines.and the windscreen surfaces, some points were added to these
The tool Multi-Section Surface of
curves. A particular in this operation regarding
the Generative Shape Design curvemodule generated was usingfinally plane
exploited to
11 is depicted in
create surfaces from the previously generated curves. Figure 7.79. A spline was created using the above mentioned
points, that is represented in Figure 7.80 (colored in red), along with the other
As mentioned before, only half of the model was recreated, therefore the surfaces in the
construction lines. The tool Multi-Section Surface in the Generative Shape Design
symmetry zone have to be tangent to the y-axis. Triangular parts near the top of the nose were
module was finally exploited to create surfaces from the previously generated
filled in suchcurves. a way to Asguarantee
mentioned tangency
before, to onlythe half
previously
of the createdmodel was surfaces.
recreated, therefore
The complete model was finally obtained
the surfaces in the symmetry zone have to be tangent to the y-axis. using symmetry. The upper and front views of the
Triangular
optimized geometry parts near arethe depicted
top of the in Figure
nose were 37. filled in such a way to guarantee tangency to
the previously created surfaces. The complete model was finally obtained using
symmetry. The upper and front views of the optimized geometry are depicted in
Copyright –Figure 2007 Agusta7.81. Once SpA,the optimized
Westland model has
Helicopters Ltd,been reconstructed,
Westland in orderLtd
Transmissions to and
evaluateAgustaWestland
Internationalthe Ltd.geometrical
(Collectivelydifferences between the optimized and the baseline configurations,
known as “AgustaWestland”)
Copyright and all other rights in this documentItare
a comparison was performed. consisted
vested ininAgustaWestland.
an overlap of the two CAD models.
The superimposition of the baseline and the optimized
This document contains confidential proprietary information and is supplied on the express condition that it may not beCAD
disclosed,models
reproduced Inis depicted
whole or in part, or in
used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
Figure 7.82.
7.24 Reverse Engineering Procedure in CATIA® V5 207
800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 61 of 98
800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 61 of 98

Figure 35: The point introduced during the curve creation in order to respect the original variation
Figure 7.79: The point introduced during
of thethe curve creation in order to respect the original
curvature.
variation of the curvature.

Figure 35: The point introduced during the curve creation in order to respect the original variation
of the curvature.

Figure 36: The cockpit construction lines.

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
 
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
Figure
Figure36: TheThe
7.80: cockpit construction
cockpit lines.
construction lines.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.
208 Part II - ERICA nose shape optimization using Ansys Fluent®

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 62 of 98

   

Figure 37: The final optimized nose geometry.


Figure 7.81: The final optimized nose geometry.
Once the optimized model has been reconstructed, in order to evaluate the geometrical
differences between the optimized and the baseline configurations, two comparisons were
performed. The first one consisted in an overlap of the two CAD models. The superimposition is
depicted in fig

800-91-067
Numerical optimisation of ERICA tiltrotor nose Rev. A
Pag. 63 of 98

Copyright – 2007 Agusta SpA, Westland Helicopters Ltd, Westland Transmissions Ltd and AgustaWestland
International Ltd. (Collectively known as “AgustaWestland”)
Copyright and all other rights in this document are vested in AgustaWestland.
This document contains confidential proprietary information and is supplied on the express condition that it may not be disclosed, reproduced In whole or in part, or used for any purpose
other than for which it is supplied, without the written consent of AgustaWestland.

   

Figura 1 - Superimposition of the baseline and the optimized CAD model.


Figure 7.82: Superimposition of the baseline and the optimized CAD models.
The second method followed to evaluate the differences between the two models is explainded
in the following. For the assumed orientation of the axes, one can refer to Figure 5.
Two families of trimming planes were built in CATIA environment, the first being constituted of
13 planes that are offset of the z-y plane, the second comprising 14 planes parallel to x-y plane.
The coordinates of these planes are presented in TABLE. For easy reference, the planes
parallel to z-y plane will be referred to as plane z-y id, where id is the identifier of the plane, the
planes parallel to x-y plane will be referred to as plane x-y id. In Table, the coordinates x (for
planes z-y id) and z (for planes x-y id) are expressed with respect to the reference point, that is the
[0 0 0].

Plan Plan Plan Plan Plane Plane Plane Plane Plane Plane Plane Plane Plane

e z-y e z-y e z-y e z-y z-y 5 z-y 6 z-y 7 z-y 8 z-y 9 z-y 10 z-y 11 z-y 12 z-y 13

1 2 3 4

1142 1312 1542 1942 2342 2742 3100 3500 3900 4300 4700 5100 5500

[mm] [mm] [mm] [mm] [mm] [mm] [mm] [mm] [mm] [mm] [mm] [mm] [mm]

Table 1 - Coordinates of planes z-y id.


Part III

Test case C:
Aerodynamic shape optimization
of the frontal region of the ERICA
tiltrotor using OpenFOAM® as
the CFD solver

209
211

7.25 Introduction
The present work deals with the ERICA tiltrotor nose shape optimization aimed
at reducing its baseline aerodynamic drag. Unlike the optimization work presented
in Part II, now a different CAD and mesh model will be addressed.
OpenFOAM® CFD open-source code was exploited to calculate aerodynamic
behavior of the model that underwent the optimization process.
Moreover, the visibility constraints were strictly taken into account via a proce-
dure described in Appendix B.
Several steps were performed in order to obtain reliable results while reducing
the overall optimization time as much as possible. First, the baseline model was
built up. To do this, the superficial mesh was first built in CATIA® environment,
and the parameterization performed with Altair Hypermesh®. Finally, the related
.hm file was built. The latter was subsequently managed in batch-mode via a script
similar to that presented in Appendix C.3.
The second step consisted in the nose aerodynamic shape optimization driven
by GeDEAII presented in Chapter 2.
As it was said above, this optimization test case featured an open-source CFD
code, that is OpenFOAM®.
Both the features and the framework of this powerful solver are presented in
Appendix D so as to make the understanding of this test case simpler.

7.26 Optimization Procedure


In this chapter, the whole sequence of operations necessary to carry out the
optimization study is described. As already mentioned, the optimization procedure
was not performed in a single step, but rather multiple steps were required to
obtain reliable results, while reducing the overall optimization time as much as
possible.
The whole sequence of operations carried out to address this work is shown in
Figure 7.83. It is worth mentioning that each block was carried out independently
from the others. The upper portion of the flow chart shows the set of preliminary
operations that need to be accomplished before starting the successive automatic
loops.
The procedure started from the elaboration of a baseline geometry CAD model
of the tiltrotor fuselage provided by AgustaWestland in CATIA® format. The
following steps focused on the construction of the baseline CFD model: first of
all, the surface mesh was generated using CATIA® surface meshing tool, then the
volume mesh was created by means of Ansys TGrid® and finally the baseline
CFD case was set up with the open-source CFD code OpenFOAM which is a
new CFD open-source code adopted at AgustaWestland for external aerodynamic
characterization of helicopter and tiltrotor fuselage components.
Once the baseline CFD solution was available, the geometry parameterization
was carried out. This operation is of outstanding importance in the optimization
212 Part III - ERICA nose shape optimization using OpenFOAM®

Geometry elaboration Sensitivity


& analysis
Surface meshing

Volume mesh

Parameterization

Optimal mesh
Parameterization
model

GeDEAII driven
optimization

Design variables Morphed surface


values mesh

CFD objective
Volume mesh
function values

Pareto front
optimal design
variables
combination

Optimal
Optimal
individuals mesh
geometries
reconstruction

Figure 7.83: Flow chart of the complete optimization procedure accommodating Open-
FOAM as the CFD colver.

process and it can be performed using various techniques. For the scope of the
present work, the commercial software HyperMesh® was chosen as the parameter-
ization tool, due to the really strong and versatile capabilities of the morphing tool
HyperMorph®.
Here some words are spent about the followed morphing strategy in this work.
In details, it was decided to parameterize the surface mesh only, and to re-mesh
every time the volume mesh, for the reasons already explained in section 7.14.
After having identified the independent variables to be taken into account, the
optimization process could be addressed, which was managed by GeDEAII as the
master program.
7.27 The object of the optimization 213

The lower part of the flow chart in Figure 7.83 describes the results extraction
and elaboration. The output of the automatic optimization loop is the optimal
combination of the design variables of the individuals lying on the Pareto frontier.
At this point, the designer can identify the individuals of interest among the
multiple optimal solutions of the Pareto front and the meshed geometry of each
selected individual can be reconstructed using HyperMorph®.
For the sake of brevity, the reverse engineering process, which allows obtaining
the deformed surface starting from the mesh, is omitted in this test case. Neverthe-
less, it can be performed following the tips already given within sections 7.12 and
7.24.
The entire optimization loop was run under Linux/UNIX environment.

7.27 The object of the optimization


The scope of this work consisted in the optimization of the ERICA nose,
whose full-scale CAD wind-tunnel model, made up of seven elements is depicted
in Figures 7.84-7.90. Only the left sides of the CAD model are depicted in the
following figures, since the simulations took into account only this part of the
ERICA model. This was possible thanks to the symmetry of the investigated flow
field.

• nose, the ERICA part that underwent the optimization process, visible in
Figure 7.84;

Figure 7.84: ERICA nose geometry in CATIA V5.

• windscreen, visible in Figure 7.85;

• fuselage, visible in Figure 7.86;

• wf-fearing, visible in Figure 7.87;


214 Part III - ERICA nose shape optimization using OpenFOAM®

Figure 7.85: ERICA windscreen geometry in CATIA V5.

Figure 7.86: ERICA fuselage geometry in CATIA V5.

• wing-sx, visible in Figure 7.88;

• sponson-sx, visible in Figure 7.89;

• tail, visible in Figure 7.90;

To this purpose, a CFD-based optimization procedure was implemented, which


was introduced in chapter 7.26, and will be described in detail in the following
chapters. The entire model was then inserted into a virtual wind tunnel, which
is made up of six elements: inlet, outlet, symmetry, sx, up e down. It is visibile in
Figure 7.91.

Box features a length of 192.8 meters, it is 58.8 [m] deep, and 117.2 [m] high.
7.27 The object of the optimization 215

Figure 7.87: ERICA wf-fearing geometry in CATIA V5.

Figure 7.88: ERICA wing-sx geometry in CATIA V5.

Figure 7.89: ERICA sponson-sx geometry in CATIA V5.


216 Part III - ERICA nose shape optimization using OpenFOAM®

Figure 7.90: ERICA tail geometry in CATIA V5.

Figure 7.91: Wind tunnel cad model.

7.28 Mesh generation


As far as the mesh generation is concerned, the softwares used to carry out the
meshing operations were respectively:

1. CATIA® surface meshing tool for the surface mesh generation;

2. Ansys Tgrid® for the volume mesh generation.

7.28.1 Superficial mesh


The CATIA® surface meshing tool can be used to mesh the geometry surfaces
directly from the CATIA® format, so no import/export operations are necessary
to accomplish this task. Moreover, the surface mesh obtained from the CATIA®
meshing tool is entirely interfaced with the CATIA® geometry. This characteristic is
important in an industrial context, because a modification of the baseline geometry
7.28 Mesh generation 217

can be directly transformed into a correspondent surface mesh modification simply


by updating the mesh.
The surface mesh was more refined over the tiltrotor nose, which was the object
of optimization, whereas the superficial grid over the cylindrical portion of the
fuselage was slightly coarser; finally, the tail cone surface mesh was the coarsest
one. The choice of adopting different mesh sizes was motivated by the attempt to
contain the total amount of surface elements (and consequently the number of 3D
cells) as much as possible.
Surface mesh is visible in Figure 7.92 where the entire wind tunnel box is
visible. Moreover, in Figure 7.93 nose and windscreen superficial mesh models are
depicted.

Figure 7.92: Superficial mesh model: wind tunnel component.

Figure 7.93: Superficial mesh model: nose component.

Some words are worth spending as regards the quality criteria accomplished
during superficial mesh generation. Actually, in order to achieve a good mesh
218 Part III - ERICA nose shape optimization using OpenFOAM®

quality it was necessary to define and set proper quality parameters before starting
the meshing operations; these parameters were then checked during the mesh
generation. AgustaWestland quality requirements can be met by imposing proper
limits on Skewness and Aspect-Ratio within CATIA®.
These parameters can be defined as follows:
• Skewness: it is a quantitative measure of the element distortion with respect
to the ideal element shape (equilateral triangle for triangular surface mesh).
Skewness is calculated within CATIA® in the following way:
S
Q = 1− (7.19)
S0
where Q is the Skewness, S is the ideal element surface, and S0 is the actual
element surface.

• Aspect-Ratio: it is the ratio between the maximum and minimum length


among the sides of the element; for instance, the Aspect-Ratio AR of a
triangular element of sides l1 , l2 , l3 is defined as:
max (l1 , l2 , l3 )
AR = (7.20)
min (l1 , l2 , l3 )

The parameters limits specified for the ERICA model superficial mesh are summa-
rized in Table 7.34. The mesh statistics of the finally selected superficial mesh with

Best value Poor at Bad at Worst value


Skewness 0 0.3 0.5 1
Aspect-Ratio 1 1.5 3 1 ∗ 106

Table 7.34: Quality parameters limits specified for ERICA surface mesh.

respect to the above mentioned criteria are summarized in Table 7.35 and Figure
7.94.

Optimal Poor ele- Bad ele- Worst ele- Mean


elements ments ments ment value
Skewness 238018 114 (0.00 %) 0 (0.00 %) 0.565 0.038
(99.95 %)
Aspect- 238132 0 (0.00 %) 0 (0.00 %) 2.34 1.157
Ratio (100 %)

Table 7.35: Mesh statistic for the finally selected ERICA superficial grid.

As apparent, AgustaWestland quality criteria were fulfilled with wide margins.


Moreover, these values are on the line with those presented in section 7.17.1.
7.28 Mesh generation 219

Figure 7.94: Histograms representing ERICA surface mesh quality statistics: Skewness (on
the left); Aspect-Ratio (on the right).

7.28.2 Volume grid


Starting from the surface mesh described above, the volume mesh was gener-
ated by means of the CFD meshing tool Ansys Tgrid®. Actually, the surface mesh
created within CATIA® can be saved in a Nastran compatible format (.dat) which
can be easily read by Tgrid®.
In Tgrid® environment, the volume between the tiltrotor surface and the wind-
tunnel walls was filled with tetrahedral elements, while the boundary layer region
was modelled using prismatic cells. In Table 7.36, the Tgrid® settings selected to
generate the internal volume mesh are presented.

Components with boundary layer nose, fuselage, fearing, tail, wing, sponson
Components without boundary layer inlet, outlet, symmetry, up, down,sx
B.L. Offset method Uniform
B.L. Growth method Geometric
B.L. Number of layer 10
B.L. First height 0.15
B.L. Growth rate 1.2
Tri/Tet Improve surface mesh option Enabled
Tri/Tet Refinment method Adv/front
Tri/Tet Cell size function Geometric
Tri/Tet Growth rate 1.2
Tri/Tet Refinement regions yes (box-fuselage, box-wing)

Table 7.36: Ansys Tgrid® settings for volume mesh generation.

Ten prismatic layers were built over the tiltrotor fuselage in order to ensure that
the simulated boundary layer was entirely contained within the prismatic layers
over the whole fuselage, with the exception of regions of separated flow. Moreover,
a height of 0.15 mm was given to the first prismatic layer in order to meet the y+
requirements for the wall function calculation during the CFD simulation ([7]).
An important aspect that has to be pointed out is the use of local refinement
220 Part III - ERICA nose shape optimization using OpenFOAM®

regions, which allowed a local improvement of the volume mesh around the ERICA
model, in particular near the nose region, while a coarser tetrahedral mesh could
be used in the remaining internal regions.
As a result, a really good mesh quality was obtained in the regions of interest
(that is, the nose and the cylindrical portion of the fuselage), with the minimum
number of 3D elements. A longitudinal view of the whole volume mesh is de-
picted in Figure 7.95, while Figure 7.96 illustrates a close-up of the mesh near the
tiltrotor surface, where the refinement regions around the fuselage can be clearly
appreciated. Finally, a particular of the boundary layer over the canopy is reported

Figure 7.95: Longitudinal view of the whole volumetric mesh around the tiltrotor fuselage.

Figure 7.96: Close-up of the volume mesh near the fuselage.

in Figure 7.97.
7.29 Fluid-dynamic model set up 221

Figure 7.97: Close-up of the volume mesh: boundary layer over the canopy.

7.29 Fluid-dynamic model set up


A pressure-based solver type with absolute velocity formulation and steady
approach was adopted for the tiltrotor simulations. The κ − ω SST model was
selected for turbulence handling. The air was treated as an ideal gas having constant
specific heats, which automatically enables the equation energy resolution. This
made it possible to include the compressibility effects in the numerical simulations:
in fact, as it can be deduced from Table 7.40, the flight condition selected for
optimization features a moderately high Mach number.
In Table 7.37, the air properties assigned for the ERICA nose optimization are
summarized. These parameters were stored into the thermophysicalProperties file
presented in Appendix D.6.3. The OpenFOAM format of this file is presented in
the following:

thermoType hPsiThermo<pureMixture<sutherlandTransport
<specieThermo<hConstThermo<perfectGas>>>>>;
mixture
{
specie
{
nMoles 1;
molWeight 28.9;
}
thermodynamics
{
Cp 1007;
Hf 0;
}
transport
{
222 Part III - ERICA nose shape optimization using OpenFOAM®

As 1.4792e-06; Ts 116;
}
}

As far as the solution algorithm is concerned, a SIMPLEC scheme was adopted,


which solves the pressure and moment equations separately. In details, the rhoSim-
plecFoam solver was addressed for this work purposes, which is a steady-state
SIMPLEC solver for laminar or turbulent RANS flow of compressible fluids. It
showed high-level performance in terms of accuracy and robustness.
The CFD calculations started with 1st order discretization schemes. The selected
1st order discretization schemes are summarized in Table 7.38, which is a excerpt of
the fvSchemes and fvSolution files used during OpenFOAM simulation. Relaxation
factors were initially given low values, in order to make more stable the iterative
search of the solution.

Gas thermo-dynamic model Ideal Gas


Cp Specific Heat [ J/kg · K ] Constant: 1007
Thermal conductivity [W/m · K ] Constant: 0.0242
Viscosity variation law Sutherland: two coefficients methods
µ0 Reference viscosity [kg/m · s] 1.4792 · e−6
T0 Reference temperature [K ] 116
S Effective temperature [K ] 110.56

Table 7.37: Air properties for ERICA nose optimization.

After 500 iterations, discretization schemes were set to a 1st − 2nd order, in
order to increase the accuracy of the solution. Moreover, relaxation factors slightly
increased to speed up the simulations.

7.29.1 Boundary and operating conditions


The boundary conditions for the optimization study were selected according to
the operating flight conditions reported in Table 7.40. In particular, turbulentIn-
tensityKineticEnergyInlet OpenFOAM boundary condition allows to calculate the
turbulent kinetic energy by means of the following relation:

3
κ= u avg I 2 (7.21)
2
that is, it is calculated from the intensity provided as a fraction of the mean
velocity. A value equal to 5% was chosen for the turbulent intensity I.
18 This boundary condition is not present in the official OpenFOAM release. As a matter of fact,

it is a home-made boundary condition which allows to calculate the turbulent frequency at the
inlet boundary by using the turbulent viscosity ratio, µt , which is the ratio between the turbulent
µ

viscosity, µt , and the molecular dynamic viscosity, µ. It proved to be an effective and robust boundary
condition.
7.29 Fluid-dynamic model set up 223

gradSchemes
{
default Gauss linear;
}
divSchemes
default Gauss upwind;
div((muEff*dev2(T(grad(U))))) Gauss linear;
}
interpolationSchemes
{
default linear;
div(U,p) linear phi;
UD linear phid;
}

relaxationFactors
{
p 0.3;
rho 0.01;
U 0.7;
h 0.7;
k 0.5;
omega 0.5;
}
relaxationFactors0
{
p 0.3;
rho 0.01;
U 0.7;
h 0.7;
k 0.5;
omega 0.5;
}

Table 7.38: 1st order discretization schemes and solver settings.

gradSchemes
{
default cellLimited Gauss linear 1;
grad(p) Gauss linear;
grad(U) Gauss linear;
grad(h) Gauss linear;
}
divSchemes
{
div(phi,U) Gauss linearUpwind grad(U);
div((muEff*dev2(T(grad(U))))) Gauss linear;
div(U,p) Gauss linearUpwind grad(p);
div(phi,h)